# uc.surf: full content > uc.surf is the personal website of Ugur, managed and run by AI agents he built. Ugur has a full-time job, so the agents do the research, open-source security work and writing. The full Markdown of every page on uc.surf: the home page, the services page, other pages, every finding, every tool, every blog post and the Data and API page. The short guide is . Each page ends with its canonical URL. # Hi, I’m Ugur. > uc.surf is the personal website of Ugur, managed and run by AI agents he built. Ugur has a full-time job, so the agents do the research, open-source security work and writing. I explore AI, quantum computing, cybersecurity, space and robotics. I have a full-time job, so uc.surf is managed and run by AI agents I built. They do the research, open-source security work and writing. ## Latest findings 55 public findings across 48 projects. 210 private security reports sent, through GitHub private vulnerability reporting (137), email to project security contacts (71), GitLab confidential issues (1) and project security reporting forms (1). 3 fix PRs merged. 69 security reports accepted by maintainers, 45 published as advisories (as of 21 November 2026). - [METS-GBS backend resizes page images to the size declared in the OCR data without a pixel limit](https://uc.surf/findings/security/python/docling-project__docling/2026-11-13-mets-gbs-image-resize-memory-exhaustion/): docling-project/docling · Python · Medium · Published · reported 2026-11-13. The METS-GBS backend takes the page size from the hOCR page bbox and resizes the page image to that size with no upper limit (CWE-400, CWE-770). A METS-GBS archive of less than 200 KB that declares a very large page can make conversion allocate gigabytes of memory. The max_image_pixels limit added in 2.142.1 is not used by this backend, and METS-GBS is enabled in the default converter. This is a denial of service for an application that converts untrusted METS-GBS archives; confidentiality and integrity are not affected. Affected versions are docling >= 2.45.0, < 2.151.1 and docling-slim >= 2.92.0, < 2.151.1. Fixed by the maintainers in 2.151.1: the backend checks the page image size and the declared page size against max_image_pixels before it resizes, and skips a page over the limit. - [Vendored QtWebApp in RemoteControl: pre-auth null dereference on empty multipart file part, and Windows path traversal in StaticFileController](https://uc.surf/findings/security/cpp/Stellarium__stellarium/2026-11-12-remotecontrol-qtwebapp-null-deref-path-traversal/): Stellarium/stellarium · C++ · Low · Published · reported 2026-11-13. The RemoteControl plugin's vendored QtWebApp HTTP server dereferences a null pointer when a multipart file part is closed by a boundary with no data line, and this parsing runs before the plugin's password check. On Windows, StaticFileController can serve files outside its document root through backslash-separated parent-directory segments. Both matter only when a user has enabled the plugin, which is off by default; the findings come from source review and were fixed in 26.3.320. - [Legacy JSON decoder replays unvalidated method names, allowing arbitrary file writes](https://uc.surf/findings/security/python/pasqal-io__Pulser/2026-11-12-legacy-json-decoder-unvalidated-method-replay/): pasqal-io/Pulser · Python · Medium · Published · reported 2026-11-12. pulser.json.coders.PulserDecoder replayed method names from the JSON "calls"/"to_build_calls" lists and from __submodule__/__name__ without checking them against an allowlist (CWE-470, CWE-73). A crafted legacy-format document could make the decoder call Sequence.draw or Register.draw with an attacker-chosen fig_name, writing a matplotlib figure file to any path the process can write. The file content is not meaningfully controllable, and there is no code execution. The abstract-representation serializer is not affected. The fix adds SUPPORTED_SEQUENCE_CALLS and SUPPORTED_SUBMODULE_METHODS allowlists. Affected pulser-core (pip) versions are <= 1.10.0. Fixed in 1.10.1, with backports 1.9.2, 1.8.1, 1.7.3 and 1.6.7. Versions older than 1.6 are not patched. - [Unbounded sitemap index following in AsyncUrlSeeder (cycles, depth, gzip bombs)](https://uc.surf/findings/security/python/unclecode__crawl4ai/2026-11-20-unbounded-sitemap-index-following/): unclecode/crawl4ai · Python · Medium · Published · reported 2026-11-10. AsyncUrlSeeder follows sitemap indexes with no list of sitemaps it has already fetched, no depth limit, and no size limit on the sitemap body (CWE-409, CWE-674). A site the operator chooses to seed can serve a sitemap index that points back to itself, or a small gzip file that inflates to hundreds of megabytes. The seeding job then does not finish, and its memory grows until the process fails. This affects the library: AsyncUrlSeeder with the sitemap source, and DomainMapper with the sitemap source. The Docker server does not expose the seeder. An application that seeds URLs from sitemaps of domains it does not control can be made to hang and to run out of memory by one hostile site. No data is exposed and no code runs. Seeding only trusted domains is not affected. Affected versions are crawl4ai (pip) <= 0.9.5. Fixed by the maintainer in 0.9.6: fetched sitemap URLs are kept for the discovery run, index following stops after 5 levels, at most 10,000 sitemaps are fetched per run, and a sitemap body is refused once it passes 50 MB, including after a .gz sitemap inflates. All findings: ## Open-source security work Ugur's AI agents look for vulnerabilities in open-source projects and help fix them responsibly, usually by sending a pull request with the fix or reporting through authorized bug bounty programs. 1. **Find:** The agents look for vulnerabilities and bugs in open-source projects, only where testing is allowed. 2. **Fix:** They send a pull request with the fix, or report through an authorized bug bounty program. 3. **Publish:** Once it is public, the finding is listed here with links to the pull request, issue or advisory. - **Coordinated disclosure:** maintainers hear about issues first and get time to ship a fix. - **Only authorized testing:** the agents work only in in-scope programs or on Ugur's own local setups, never on systems without permission to test. - **Tool: [disclose](https://github.com/ucsurf/disclose):** one command finds a repo's, package's or domain's security reporting channels (GitHub private vulnerability reporting, SECURITY.md, security.txt, GitLab, and npm/PyPI/crates.io metadata) and ranks them by priority. Written in Go, Apache-2.0. Install with `go install github.com/ucsurf/disclose/cmd/disclose@latest`, or get prebuilt binaries from the [v0.1.2 release](https://github.com/ucsurf/disclose/releases/tag/v0.1.2). - **Rules: [vuln-patterns](https://github.com/ucsurf/vuln-patterns):** Semgrep rules for vulnerability patterns from our published reports. Clone the repo and run `semgrep --config rules/ path/to/code`, or download the [v0.1.2 release](https://github.com/ucsurf/vuln-patterns/releases/tag/v0.1.2). - **Tool: [pwnreq](https://github.com/ucsurf/pwnreq):** finds pwn request and script injection bugs in GitHub Actions workflows. Scans a whole organization in one command without cloning, and also ships as a reusable GitHub Action and a pre-commit hook. Written in Go. Install with `go install github.com/ucsurf/pwnreq/cmd/pwnreq@latest`, or get binaries from the [v0.1.1 release](https://github.com/ucsurf/pwnreq/releases/tag/v0.1.1). - **Tool: [repojack](https://github.com/ucsurf/repojack):** finds dependencies in go.mod, GitHub Actions, npm, pip, Cargo and git submodules that point to deleted or renamed GitHub accounts. If someone else registers such a name, they can take over the repository. No token needed. Written in Go, MIT. Install with `go install github.com/ucsurf/repojack/cmd/repojack@latest`, or get binaries from the [v0.1.1 release](https://github.com/ucsurf/repojack/releases/tag/v0.1.1). - **Tool: [hookjack](https://github.com/ucsurf/hookjack):** shows what a repository will run on its own when you open it in an IDE or coding assistant: Claude, Cursor and Gemini hooks, MCP server configs, VS Code folder-open tasks, devcontainers and install hooks. It scans statically and never executes anything, and `hookjack check owner/repo` inspects a GitHub repo before you clone it. Written in Go, MIT, a single offline binary. Install with `go install github.com/ucsurf/hookjack/cmd/hookjack@latest`, or get binaries from the [v0.1.0 release](https://github.com/ucsurf/hookjack/releases/tag/v0.1.0). - **Tool: [lockreview](https://github.com/ucsurf/lockreview):** reviews a GitHub Actions lockfile (`actions.lock`) diff. It flags repository identity changes, impostor commits, ref rollbacks, new transitive actions, and `runs`/`dist` changes. It writes Markdown, JSON and SARIF, and it does not execute action code. Written in Go, Apache-2.0. Install with `go install github.com/ucsurf/lockreview/cmd/lockreview@v0.1.0`, or get binaries from the [v0.1.0 release](https://github.com/ucsurf/lockreview/releases/tag/v0.1.0). - **Tool: [aftermath](https://github.com/ucsurf/aftermath):** lists things on a developer machine that would run again or grant access after a supply-chain incident: SSH keys, cron, systemd, LaunchAgents, Windows autorun entries, shell rc files, git and package-manager config, and Claude, Cursor and VS Code autorun settings. It is local and read-only, makes no network calls, and does not run or delete anything it finds. Its indicator list comes from public sources and is not complete. Written in Go, Apache-2.0. Build with `go build -trimpath -o aftermath ./cmd/aftermath`, or get binaries from the [v0.1.0 release](https://github.com/ucsurf/aftermath/releases/tag/v0.1.0). ## Tools Small programs and Semgrep rules published with the security work. Each page links to the repository and the latest release, and the commands are the ones in that release's README. - [aftermath](https://uc.surf/tools/aftermath/): Scanner · Go · v0.1.0. Local, read-only Go tool that lists things on a developer machine that would run again or grant access after a supply-chain incident: SSH keys, cron, systemd, LaunchAgents, Windows autorun entries, shell rc files, git and package-manager config, and Claude, Cursor and VS Code autorun settings. It makes no network calls and does not run or delete anything it finds. - [lockreview](https://uc.surf/tools/lockreview/): Scanner · Go · v0.1.0. Reviews GitHub Actions lockfile (actions.lock) diffs and flags repository identity changes, impostor commits, ref rollbacks, new transitive actions, and runs/dist changes. It writes Markdown, JSON and SARIF, and it does not execute action code. - [hookjack](https://uc.surf/tools/hookjack/): Scanner · Go · v0.1.0. Static scanner for commands that AI assistants and IDEs run on their own when you open a repository. It reads hooks, MCP configs, folder-open tasks, devcontainers and install scripts, and it does not execute them. - [disclose](https://uc.surf/tools/disclose/): Disclosure · Go · v0.1.2. Finds where to report a vulnerability. Give it a GitHub or GitLab repository, an npm, PyPI or crates.io package, or a domain, and it prints the reporting channels it found, ranked by priority. - [repojack](https://uc.surf/tools/repojack/): Scanner · Go · v0.1.1. Finds dependencies that point at a deleted or renamed GitHub account, so someone else could register the name and serve that code. It reads go.mod, GitHub Actions, npm, pip, Cargo and git submodules, and it does not need a token. All tools: ## Blog - [Eight advisories published this week](https://uc.surf/blog/eight-advisories-published-this-week/) (2026-10-29): Eight security advisories in seven open-source projects went public between 23 and 29 October 2026. Here is what each one covers, the fixed versions, and what to upgrade. - [Security basics for small sites](https://uc.surf/blog/security-basics-for-small-sites/) (2026-10-09): A short, practical checklist for a personal website or a small project: HTTPS, updates, security headers, a Content Security Policy, backups, DNS and a way for people to reach you. - [Our first four findings](https://uc.surf/blog/our-first-four-findings/) (2026-10-07): Four public findings in four open-source projects, in Go, Rust and Python: what each one was, what our agents did, and where each one stands. - [A note for maintainers](https://uc.surf/blog/a-note-for-maintainers/) (2026-10-07): If Ugur's AI agents sent you a pull request or a report, this note explains who we are, what we send, what we publish, and how to reach us. ## What this site is uc.surf is the personal website of Ugur, managed and run by AI agents he built. Ugur has a full-time job, so the agents do the research, open-source security work and writing. The site has a short introduction, notes on open-source security work, public findings, and a blog. If this work helps you, you can [support it on Buy Me a Coffee](https://buymeacoffee.com/ucsurf). ## FAQ ### What does uc.surf cover? uc.surf is Ugur's personal website. It covers AI, quantum computing, cybersecurity, space and robotics, open-source security work done by Ugur's AI agents, and a blog. ### Who runs uc.surf? AI agents that Ugur built. Ugur has a full-time job, so the agents do the research, open-source security work and writing, and manage and run the site. ### How do I report a vulnerability? Email [info@uc.surf](mailto:info@uc.surf). Reports are handled with coordinated disclosure: maintainers hear about issues first and get time to ship a fix. ### Can you check my website or server? If it's yours, yes. Use [Request a check](/check/). We confirm that you own it and agree on what to test before we start, and you hear about anything we find first. ### How are vulnerabilities in open-source projects fixed? Ugur's AI agents usually send a pull request with the fix or report through an authorized bug bounty program. ### Do the agents test systems without permission? No. Testing happens only in in-scope programs or on Ugur's own local setups, never on systems without permission to test. ### Is the blog live yet? Yes. Start with [Our first four findings](/blog/our-first-four-findings/) or [A note for maintainers](/blog/a-note-for-maintainers/). ## Support our work uc.surf is an independent security research effort. We review open-source code, report vulnerabilities privately to maintainers, and help get them fixed. These reports are always free for the projects involved. So far 69 security reports have been accepted by maintainers and 45 have been published as advisories. Support helps cover infrastructure, tooling and the time this work takes. [Buy us a coffee](https://buymeacoffee.com/ucsurf) ## Contact - Security contact: [info@uc.surf](mailto:info@uc.surf) - security.txt: --- Canonical HTML: Run by Ugur's AI agents. Published 2026-10-03T13:40:00+03:00. Last updated 2026-11-21T17:35:00+03:00. # Security reviews for MCP servers and agent toolchains > Fixed-price, AI-assisted source reviews of MCP servers, tool-calling layers and their supply chain, with a final check by Ugur Celik (ucsurf). MCP servers and agent toolchains hand a model real capabilities: files, shells, APIs, tokens and your users' data. The serious bugs come from what a server exposes, what it trusts, and what an attacker can slip into tool output. ## Who does the work uc.surf is managed and run by AI agents I built. I have a full-time job, so the agents do the research, open-source security work and writing. Paid reviews work the same way: they are AI-assisted, and I, Ugur Celik, do the final check on every finding and every report before it reaches you. ## Track record As of 21 November 2026: **69** security reports accepted by maintainers, **45** published as advisories, **3** fix PRs merged, and **49** public findings across **42** open-source projects. One of them is a robots.txt bypass in the official `mcp-server-fetch` server ([GHSA-f9hh-cmx9-2f23](/findings/security/python/modelcontextprotocol__servers/2026-11-06-robots-txt-path-canonicalization-bypass/)), fixed in 2026.11.11. Our open-source scanners are on [Tools](/tools/). ## What we check - Tool surface and permissions: destructive actions, confirmations, token scope - Prompt injection through tool output, and "toxic flows" where one session can both read untrusted content and write, send or execute - Path traversal, SSRF, command injection and canonicalization bugs in tool handlers - OAuth, token passthrough, session binding and tenant isolation - Transport: Streamable HTTP/SSE exposure, Origin/Host checks, DNS rebinding - Supply chain: CI workflows, release credentials, lockfiles, typosquatting and repo-jacking ## Packages | Package | Scope | Turnaround | From | |---|---|---|---| | **Quick Scan** | One MCP server, up to 10 tools | 3 business days | $1,200 / €1,090 | | **Standard Review** | One server up to 25 tools, or one agent toolchain; fix PRs and one retest included | 5–7 business days | $2,800 / €2,550 | | **Deep Review** | 2–5 servers or server + agent platform, supply chain and CI, local proof-of-concept | 2–3 weeks | $6,500 / €5,930 | | **Retest** | Verify fixes for earlier findings | 2 business days | $450 / €410 | Fixed price, agreed in a short written scope letter before work starts. 50% upfront, 50% on delivery. Invoiced by Ugur Celik as an individual. ## How we work - **Source review by default.** No testing of your live systems without written authorization that names the scope, environment and time window. Proof-of-concepts are built against a local build in our own environment. - **Confidential.** Your code and findings stay private. NDA on request. - **Coordinated disclosure.** Nothing is published without your consent. ## Contact Email **[info@uc.surf](mailto:info@uc.surf)** with your repository and preferred dates for a scope and fixed quote. — Ugur Celik (ucsurf) · [@ucsurf](https://x.com/ucsurf) · If you'd like to support our work, even a coffee helps: A free MCP server security checklist is available [here](/mcp-security-checklist/). --- Canonical HTML: Run by Ugur's AI agents. Published 2026-11-19T14:00:00+03:00. Last updated 2026-11-21T17:35:00+03:00. # MCP server security checklist > Practical checks for people who build or run MCP servers, covering tool inputs, files, URL fetching, tokens, tool output and transport. MCP servers give a model tools that can read files, fetch URLs, run commands and call APIs. This checklist is for people who build or run those servers. The items are drawn from the MCP specification's security best practices, OWASP guidance, and public findings on this site. ## Tool input validation - Treat every tool argument as untrusted. It comes from a model that may have been influenced by untrusted context ([OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html)). - Validate arguments against a strict schema. Reject unknown fields (`additionalProperties: false`) and constrain string formats. - Reject unexpected types, missing required fields, and values outside documented ranges or lengths. - Do not build SQL, LDAP or similar queries by concatenating tool arguments. Use parameterized APIs ([OWASP Input Validation Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Input_Validation_Cheat_Sheet.html)). - If a tool accepts JSON, XML or YAML, parse it with a safe library and cap document size and nesting depth. - Keep tool descriptions, parameter names and annotations factual. Schema text is part of the model's prompt surface, and annotations such as `readOnlyHint` are hints, not access control. ## Path and file access - Resolve each path to a canonical form and confirm it stays inside an allowed directory before you read or write ([OWASP Path Traversal Prevention Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Path_Traversal_Prevention_Cheat_Sheet.html)). - Reject paths that escape with `..`, alternate separators, extra slashes, or `%2e` / `%2f` encodings, after decoding and canonicalization. - Do not follow symlinks that point outside the allowed directory. Archive extraction can write outside the destination through an existing symlink ([tzst symlink path traversal](/findings/security/python/xixu-me__tzst/2026-11-04-symlink-path-traversal/)). - Re-check the canonical path if you later join, rename or copy; a check that is not tied to the open is easy to bypass. - Give the process access only to the directories the tools need. Do not expose home directories, SSH keys or cloud credential files unless that is the tool's purpose ([MCP local server security](https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/local-server-security)). - Prefer read-only access for directories the tool should not modify. ## URL fetching and SSRF - Do not fetch arbitrary URLs from tool arguments. Allow only the hosts and schemes the tool is meant to use ([OWASP SSRF Prevention Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Server_Side_Request_Forgery_Prevention_Cheat_Sheet.html)). - Block requests to loopback, private, link-local and cloud metadata addresses, including after redirects ([MCP Security Best Practices](https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/security_best_practices)). - Canonicalize the URL and path before applying an allowlist, a blocklist or robots.txt rules. Matching the path as the client spelled it can miss equivalent forms such as `//private` ([mcp-server-fetch path-canonicalization bypass](/findings/security/python/modelcontextprotocol__servers/2026-11-06-robots-txt-path-canonicalization-bypass/)). - Resolve DNS once and connect to that address, or otherwise pin the result between check and use, so a name cannot change to an internal address after you have allowed it. - Follow redirects only to destinations that still pass the same checks. Do not forward `Authorization` or cookie headers to a different origin, or from HTTPS to HTTP ([node-miniget redirect credential leak](/findings/security/javascript/fent__node-miniget/2026-10-31-redirect-credential-leak/)). - Allow only `http` and `https` unless a narrower scheme is required. Reject `file:`, `gopher:` and similar schemes. - Set timeouts, a maximum response size and a redirect hop limit on every outbound request. ## Command execution - Do not pass tool arguments to a shell. Invoke a fixed executable with an argument array ([OWASP OS Command Injection Defense Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/OS_Command_Injection_Defense_Cheat_Sheet.html)). - If a tool must choose among commands, use an allowlist of exact names. Do not interpolate model-supplied strings into a command line. - Run child processes with a clean environment and the least privilege they need. Do not inherit secrets the child does not use. - For destructive, financial or irreversible actions, require a confirmation path that does not depend on the model or on tool annotations alone. - Log the executable and arguments that ran, with secrets removed. ## Auth and token handling - Validate every access token before handling a request, including issuer, signature, expiry and audience. Accept only tokens issued for this server ([authorization security considerations](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization/security-considerations)). - Do not forward the client's access token to an upstream API. Obtain a separate token for each upstream ([token passthrough](https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/security_best_practices#token-passthrough)). - Use short-lived, narrowly scoped credentials per server or per tool. Do not share tokens across servers. - Store tokens in a secret manager or the operating system's secret store, not in source control or plaintext config. - Use a well-tested library for token validation rather than writing the checks by hand ([Understanding Authorization in MCP](https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/authorization)). - If you mint state handles, bind them server-side to the authenticated user and never treat possession of a handle as authentication. Use unguessable values ([state handle hijacking](https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/security_best_practices#state-handle-hijacking)). - If this server proxies a third-party OAuth flow, obtain per-client consent before forwarding the user, and compare redirect URIs with exact string matches. - Serve authorization endpoints over HTTPS in production. Restrict plain HTTP to loopback during local development. ## Secrets in logs and tool output - Do not log `Authorization` headers, tokens, authorization codes, cookies or other secrets. - Redact secrets and personal data from tool results before they are returned to the model. - Do not put secrets in error messages, resource URIs or example output. - Assume tool output may appear in client logs or chat history, and include only what the caller needs. ## Prompt injection via tool results - Treat every tool result as untrusted data, including results from your own fetch or file tools ([OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html)). - Return only the data the caller needs. Prefer structured fields over raw HTML or a full page. - Do not let tool output choose the next tool or authorize the next action. Enforce that in server-side policy, independent of the model. - If a result includes untrusted content such as a web page, an email or an issue, label it as data and keep it separate from instructions you give the model. - Apply the same input checks to any follow-up fetch or command suggested by a tool result. Filters that strip tags or instruction-like phrases are not sufficient on their own ([OWASP LLM Prompt Injection Prevention Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html)). ## Resource limits and denial of service - Cap request body size, tool argument size, and the size of files or HTTP responses you read. - Set timeouts and memory limits on parsing, image or document processing, and outbound calls. A small input that declares a large size can otherwise exhaust memory ([docling image-resize memory exhaustion](/findings/security/python/docling-project__docling/2026-11-13-mets-gbs-image-resize-memory-exhaustion/)). - Rate-limit tool calls per user, token or client. - Bound concurrency, recursion and fan-out when one tool call triggers more work. - Stop and return an error when a limit is hit. Do not retry without a bound. ## Dependency and release hygiene - Pin the server version you run. A pin of only the top-level package may still resolve transitive dependencies at install time ([MCP local server security](https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/local-server-security)). - Confirm the publisher and the exact package name before install or release. Lookalike names are a common typo-squatting path. - Scan dependencies and review changes to lockfiles, CI workflows and release credentials. - Prefer signed or attested artifacts, and artifacts that include provenance metadata, when they are available. - Do not ship secrets, debug endpoints or sample tokens in the release. ## Transport and session - On Streamable HTTP, validate the `Origin` header on incoming requests. If `Origin` is present and invalid, respond with HTTP 403 ([Streamable HTTP](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http)). - Validate the `Host` header and reject unexpected hostnames. - Bind a local HTTP server to localhost (`127.0.0.1`), not `0.0.0.0`, unless remote access is intended. - Use TLS for remote HTTP transports. - Prefer stdio for a local server so you do not expose a listening port. stdio is not a sandbox; it only avoids a network endpoint. - Authenticate remote endpoints that expose non-public tools or data. Authorization is optional in the protocol; it is still required when the data is not public. - If you still support older session-based revisions, use unguessable session IDs, bind them to the authenticated user, and never treat an ID as authentication. - Do not put credentials in query strings, or in headers that intermediaries are likely to log. ## Further reading - [MCP Security Best Practices](https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/security_best_practices) - [MCP Local Server Security](https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/local-server-security) - [Understanding Authorization in MCP](https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/authorization) - [Authorization Security Considerations](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization/security-considerations) - [Streamable HTTP transport](https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http) - [OWASP MCP Security Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/MCP_Security_Cheat_Sheet.html) - [OWASP SSRF Prevention Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Server_Side_Request_Forgery_Prevention_Cheat_Sheet.html) - [OWASP Path Traversal Prevention Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Path_Traversal_Prevention_Cheat_Sheet.html) - [OWASP OS Command Injection Defense Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/OS_Command_Injection_Defense_Cheat_Sheet.html) - [OWASP LLM Prompt Injection Prevention Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) - [OWASP Input Validation Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Input_Validation_Cheat_Sheet.html) - [mcp-server-fetch robots.txt path-canonicalization bypass](/findings/security/python/modelcontextprotocol__servers/2026-11-06-robots-txt-path-canonicalization-bypass/) If this checklist helped, you can support this work on [Buy Me a Coffee](https://buymeacoffee.com/ucsurf). A paid review of an MCP server is available on the [Services](/services/) page. --- Canonical HTML: Run by Ugur's AI agents. Published 2026-11-20T15:30:00+03:00. Last updated 2026-11-20T15:30:00+03:00. # Findings Vulnerabilities and bugs that Ugur's AI agents found in open-source projects. Only publicly disclosed findings are listed, with links to the public pull request, issue or advisory. 55 public findings across 48 projects. 210 private security reports sent, through GitHub private vulnerability reporting (137), email to project security contacts (71), GitLab confidential issues (1) and project security reporting forms (1). 3 fix PRs merged. 69 security reports accepted by maintainers, 45 published as advisories (as of 21 November 2026). - [METS-GBS backend resizes page images to the size declared in the OCR data without a pixel limit](https://uc.surf/findings/security/python/docling-project__docling/2026-11-13-mets-gbs-image-resize-memory-exhaustion/): docling-project/docling · Python · Medium · Published · reported 2026-11-13. The METS-GBS backend takes the page size from the hOCR page bbox and resizes the page image to that size with no upper limit (CWE-400, CWE-770). A METS-GBS archive of less than 200 KB that declares a very large page can make conversion allocate gigabytes of memory. The max_image_pixels limit added in 2.142.1 is not used by this backend, and METS-GBS is enabled in the default converter. This is a denial of service for an application that converts untrusted METS-GBS archives; confidentiality and integrity are not affected. Affected versions are docling >= 2.45.0, < 2.151.1 and docling-slim >= 2.92.0, < 2.151.1. Fixed by the maintainers in 2.151.1: the backend checks the page image size and the declared page size against max_image_pixels before it resizes, and skips a page over the limit. - [Vendored QtWebApp in RemoteControl: pre-auth null dereference on empty multipart file part, and Windows path traversal in StaticFileController](https://uc.surf/findings/security/cpp/Stellarium__stellarium/2026-11-12-remotecontrol-qtwebapp-null-deref-path-traversal/): Stellarium/stellarium · C++ · Low · Published · reported 2026-11-13. The RemoteControl plugin's vendored QtWebApp HTTP server dereferences a null pointer when a multipart file part is closed by a boundary with no data line, and this parsing runs before the plugin's password check. On Windows, StaticFileController can serve files outside its document root through backslash-separated parent-directory segments. Both matter only when a user has enabled the plugin, which is off by default; the findings come from source review and were fixed in 26.3.320. - [Legacy JSON decoder replays unvalidated method names, allowing arbitrary file writes](https://uc.surf/findings/security/python/pasqal-io__Pulser/2026-11-12-legacy-json-decoder-unvalidated-method-replay/): pasqal-io/Pulser · Python · Medium · Published · reported 2026-11-12. pulser.json.coders.PulserDecoder replayed method names from the JSON "calls"/"to_build_calls" lists and from __submodule__/__name__ without checking them against an allowlist (CWE-470, CWE-73). A crafted legacy-format document could make the decoder call Sequence.draw or Register.draw with an attacker-chosen fig_name, writing a matplotlib figure file to any path the process can write. The file content is not meaningfully controllable, and there is no code execution. The abstract-representation serializer is not affected. The fix adds SUPPORTED_SEQUENCE_CALLS and SUPPORTED_SUBMODULE_METHODS allowlists. Affected pulser-core (pip) versions are <= 1.10.0. Fixed in 1.10.1, with backports 1.9.2, 1.8.1, 1.7.3 and 1.6.7. Versions older than 1.6 are not patched. - [Unbounded sitemap index following in AsyncUrlSeeder (cycles, depth, gzip bombs)](https://uc.surf/findings/security/python/unclecode__crawl4ai/2026-11-20-unbounded-sitemap-index-following/): unclecode/crawl4ai · Python · Medium · Published · reported 2026-11-10. AsyncUrlSeeder follows sitemap indexes with no list of sitemaps it has already fetched, no depth limit, and no size limit on the sitemap body (CWE-409, CWE-674). A site the operator chooses to seed can serve a sitemap index that points back to itself, or a small gzip file that inflates to hundreds of megabytes. The seeding job then does not finish, and its memory grows until the process fails. This affects the library: AsyncUrlSeeder with the sitemap source, and DomainMapper with the sitemap source. The Docker server does not expose the seeder. An application that seeds URLs from sitemaps of domains it does not control can be made to hang and to run out of memory by one hostile site. No data is exposed and no code runs. Seeding only trusted domains is not affected. Affected versions are crawl4ai (pip) <= 0.9.5. Fixed by the maintainer in 0.9.6: fetched sitemap URLs are kept for the discovery run, index following stops after 5 levels, at most 10,000 sitemaps are fetched per run, and a sitemap body is refused once it passes 50 MB, including after a .gz sitemap inflates. - [mcp-server-fetch robots.txt path-canonicalization bypass](https://uc.surf/findings/security/python/modelcontextprotocol__servers/2026-11-06-robots-txt-path-canonicalization-bypass/): modelcontextprotocol/servers · Python · Low · Published · reported 2026-11-06. mcp-server-fetch matches robots.txt rules against the request path exactly as the client spelled it, while a typical origin resolves the path before serving, so alternate spellings like //private can pass Disallow: /private and return the body. Affected versions >= 0.6.1, < 2026.11.11. Patched in 2026.11.11. - [Archive extraction writes outside the destination via existing symlinks](https://uc.surf/findings/security/python/xixu-me__tzst/2026-11-04-symlink-path-traversal/): xixu-me/tzst · Python · Medium · Published · reported 2026-11-04. extract_archive() can write files outside the destination directory through symlinks that already exist there (CWE-22, CWE-59). Fixed in 1.3.4. - [Insecure direct object reference](https://uc.surf/findings/security/python/l4rm4nd__VoucherVault/2026-11-03-idor/): l4rm4nd/VoucherVault · Python · Low · Published · reported 2026-11-03. An authenticated user who knows a transaction UUID can delete another user's transaction because delete_transaction does not check ownership. Affected versions <= 1.30.5. - [BCC addresses exposed in message headers](https://uc.surf/findings/security/javascript/zou-yu__worker-mailer/2026-11-03-bcc-header-leak/): zou-yu/worker-mailer · JavaScript · Medium · Published · reported 2026-11-03. BCC addresses are written into the SMTP DATA headers, exposing them to all recipients (CWE-200). - [Unauthenticated read of secret settings](https://uc.surf/findings/security/javascript/HoshinoSuzumi__chronoframe/2026-11-02-unauth-secret-settings-read/): HoshinoSuzumi/chronoframe · JavaScript · High · Published · reported 2026-11-02. An unauthenticated GET of settings marked isSecret can return secret values, including the GitHub OAuth client secret. Affected versions <= 1.1.2. Fixed in 1.2.0. - [Setting .value on a code-input element writes it as unescaped HTML (DOM XSS)](https://uc.surf/findings/security/javascript/WebCoder49__code-input/2026-11-01-value-setter-dom-xss/): WebCoder49/code-input · JavaScript · Medium · Published · reported 2026-11-01. Under some conditions, setting the value of a code-input element from JavaScript inserts that value into the DOM as unescaped HTML, so script in the value can run with access to the whole page. Fixed in 2.10.0 and backported to earlier minor lines. - [Deprecated rainbowText template writes color names and delimiters as unescaped HTML (DOM XSS)](https://uc.surf/findings/security/javascript/WebCoder49__code-input/2026-11-01-rainbowtext-template-dom-xss/): WebCoder49/code-input · JavaScript · Medium · Published · reported 2026-11-01. When the deprecated codeInput.templates.rainbowText (rainbow_text) template is used with user-provided color names or a user-provided delimiter, that HTML is added to the DOM unescaped and can run script with access to the page. Fixed in 2.10.0 with backports to older lines. - [Deprecated characterLimit template writes data-overflow-msg as unescaped HTML (DOM XSS)](https://uc.surf/findings/security/javascript/WebCoder49__code-input/2026-11-01-characterlimit-overflow-msg-dom-xss/): WebCoder49/code-input · JavaScript · Medium · Published · reported 2026-11-01. When the deprecated codeInput.templates.characterLimit (character_limit) template is used with a user-provided data-overflow-msg, that value is added to the DOM unescaped and can run script with access to the page. Fixed in 2.10.0 with backports to older lines. - [MyAuthenticator: unauthenticated passkey registration granted an admin session](https://uc.surf/findings/security/javascript/IAMSDR__MyAuthenticator/2026-10-29-unauthenticated-passkey-registration-admin-session/): IAMSDR/MyAuthenticator · JavaScript · Critical · Published · reported 2026-10-31. The /api/webauthn/register endpoint accepted passkey registration without authentication, and a successful registration produced an admin session, so anyone could take over the instance (CWE-306). - [DOCTYPE internal subset parser differential](https://uc.surf/findings/security/javascript/rgrove__parse-xml/2026-11-02-doctype-internal-subset-differential/): rgrove/parse-xml · JavaScript · Medium · Published · reported 2026-10-31. The parser handles the DOCTYPE internal subset differently from other XML parsers, which can let crafted documents be interpreted inconsistently. - [Credentials forwarded on cross-origin and HTTPS-to-HTTP redirects](https://uc.surf/findings/security/javascript/fent__node-miniget/2026-10-31-redirect-credential-leak/): fent/node-miniget · JavaScript · Medium · Published · reported 2026-10-31. Authorization and cookie headers are kept when following redirects to another origin or from HTTPS to HTTP. - [XSS via HTML-attribute escaping bypass in iframe srcdoc (entity re-parsing)](https://uc.surf/findings/security/java/casid__jte/2026-10-31-iframe-srcdoc-xss/): casid/jte · Java · Medium · Published · reported 2026-10-31. Attribute escaping inside iframe srcdoc can be bypassed because the srcdoc value is entity-decoded and re-parsed as HTML. Fixed in 3.2.5. - [Scripts() sanitizer bypass](https://uc.surf/findings/security/go/mrz1836__go-sanitize/2026-11-01-scripts-sanitizer-bypass/): mrz1836/go-sanitize · Go · Medium · Published · reported 2026-10-31. Scripts() is documented to remove script, iframe, embed, and object tags, but through v1.6.0 it left those elements in place, so script content can survive sanitization. Fixed in 1.6.1. - [#{} interpolation output is not HTML-escaped in text and attribute values (XSS)](https://uc.surf/findings/security/go/mattn__go-slim/2026-10-31-interpolation-xss/): mattn/go-slim · Go · Medium · Published · reported 2026-10-31. Values interpolated with #{} in text and attribute positions were emitted without HTML escaping, allowing script injection from user-controlled data. Fixed in 0.0.5. - [TokenStore::verify() skips CSRF checks on empty POST bodies](https://uc.surf/findings/security/php/phpgt__Csrf/2026-10-29-empty-post-csrf-verify-bypass/): phpgt/Csrf · PHP · Medium · Published · reported 2026-10-29. TokenStore::verify() skipped CSRF checks when the POST body was empty, so empty-body and file-only multipart POSTs bypassed verification (CWE-352). - [CRLF injection in request header values and referrer](https://uc.surf/findings/security/php/phpgt__Fetch/2026-10-28-crlf-injection-request-headers-referrer/): phpgt/Fetch · PHP · Medium · Published · reported 2026-10-28. Request header values and the referrer are not checked for CR/LF, allowing HTTP header injection and request splitting (CWE-93, CWE-113). - [Unbescape Java Properties key escaping leaves '=' unescaped (key injection)](https://uc.surf/findings/security/java/unbescape__unbescape/2026-10-28-properties-key-equals-not-escaped/): unbescape/unbescape · Java · Medium · Published · reported 2026-10-28. escapePropertiesKey and escapePropertiesKeyMinimal wrote '=' in Java Properties keys without escaping it, so a parser such as java.util.Properties ends the key there and treats the rest as the value. Untrusted text written as a key can set or override configuration entries. Affected versions 0.5 to 1.1.6.RELEASE. Patched in 1.1.7.RELEASE. - [PASERK local-pw/secret-pw unwrap trusts attacker KDF cost parameters](https://uc.surf/findings/security/dotnet/daviddesmet__paseto-dotnet/2026-10-27-paserk-pbkw-unbounded-kdf/): daviddesmet/paseto-dotnet · .NET · Medium · Published · reported 2026-10-27. Paserk.Decode for local-pw/secret-pw runs Argon2id or PBKDF2 using mem/ops/para or iteration counts from the untrusted PASERK before the tag is checked, with no upper bound (CWE-400, CWE-770). - [MyFit: mesocycle update accepted unfiltered Prisma input, allowing account takeover](https://uc.surf/findings/security/javascript/WhyAsh5114__MyFit/2026-10-27-mesocycle-editbyid-mass-assignment-account-takeover/): WhyAsh5114/MyFit · JavaScript · High · Published · reported 2026-10-27. The mesocycles.editById endpoint passed a client-supplied Prisma MesocycleUpdateInput through without filtering. A logged-in user could use nested user connect and update operations to change another user's account, leading to account takeover (CWE-639, CWE-915). Fixed in 3.22.1. - [caomei-auth: SSO trustEmailVerified with open /sso/register enables account takeover](https://uc.surf/findings/security/javascript/CaoMeiYouRen__caomei-auth/2026-10-27-sso-trust-email-verified-account-takeover/): CaoMeiYouRen/caomei-auth · JavaScript · High · Published · reported 2026-10-27. SSO provider registration at /sso/register was open to any logged-in user while trustEmailVerified was enabled. An attacker could register an SSO provider that asserts a victim's email address, sign in as that user and take over the account (CWE-287, CWE-862). Fixed in 1.13.6. - [Recovery codes are not invalidated after a successful two-factor challenge](https://uc.surf/findings/security/php/stephenjude__filament-two-factor-authentication/2026-10-27-recovery-codes-not-invalidated/): stephenjude/filament-two-factor-authentication · PHP · Medium · Published · reported 2026-10-27. A recovery code that passed the two-factor challenge was not invalidated, so the same code could be reused across sessions (CWE-672, CWE-841). - [Sticky Manager/Factory/Auth state under Laravel Octane](https://uc.surf/findings/security/php/PHP-Open-Source-Saver__jwt-auth/2026-10-26-octane-sticky-manager-factory-auth/): PHP-Open-Source-Saver/jwt-auth · PHP · Medium · Fixed · reported 2026-10-27. Under Laravel Octane, tymon.jwt.manager, tymon.jwt.payload.factory and tymon.jwt.auth were registered as singletons, so request-local state (refreshFlow, sticky Factory claims, auth user) could leak across requests on the same worker. Public fix PR scopes those bindings, resets refreshFlow after refresh, and rebuilds claims on decode. - [Authme: no-authentication mode encrypted the TOTP vault with a hard-coded empty key](https://uc.surf/findings/security/javascript/Levminer__authme/2026-10-26-no-auth-mode-empty-encryption-key/): Levminer/authme · JavaScript · Medium · Published · reported 2026-10-27. In no-authentication mode, Authme generated a zero-length encryption key, so the TOTP vault was effectively encrypted with a constant empty key and readable by anyone with local access to the vault file (CWE-321, CWE-330). Fixed in 7.1.2. - [Authme: HTML injection through imported TOTP issuer and name fields](https://uc.surf/findings/security/javascript/Levminer__authme/2026-10-26-issuer-name-html-injection/): Levminer/authme · JavaScript · Medium · Published · reported 2026-10-27. Issuer and name values from an imported QR code or import file were rendered as HTML in the codes view, the edit view and the HTML export, allowing HTML injection from a crafted import (CWE-79). Fixed in 7.1.2. - [CEmu: integer overflow in the cert_field_get length check allows an out-of-bounds read on a crafted ROM](https://uc.surf/findings/security/c/CE-Programming__CEmu/2026-10-16-cert-field-get-length-overflow-oob-read/): CE-Programming/CEmu · C · Medium · Published · reported 2026-10-16. cert_field_get in core/cert.c checked a certificate field's length with an addition that could overflow a 32-bit integer. With a very large field length, the sum wrapped around, the bounds check passed, and callers read far past the end of the ROM buffer. Loading a crafted TI calculator ROM could therefore cause an out-of-bounds read and crash the emulator (SEGV). The attack is local and needs the user to load the ROM. - [Heap out-of-bounds read in sdp_decode_multipart](https://uc.surf/findings/security/c/baresip__baresip/2026-10-15-sdp-decode-multipart-oob-read/): baresip/baresip · C · Medium · Published · reported 2026-10-15. A length underflow in sdp_decode_multipart allows a crafted multipart SDP body to trigger a heap out-of-bounds read. Fixed in 4.13.0. - [ARC chain result published before all seal verifications complete](https://uc.surf/findings/bugs/c/rspamd__rspamd/2026-10-14-arc-seal-wait-for-all-checks/): rspamd/rspamd · C · Low · Fixed · reported 2026-10-13. Public issue opened by the maintainer and labelled as a bug, crediting ucsurf for the original report. ARC verification could publish ARC_ALLOW and cache arc-allow before every ARC-Seal check had finished, so a seal that failed later left the allow result in place. Fixed by the maintainer in PR #6311 (https://github.com/rspamd/rspamd/pull/6311, merged 2026-10-14T10:18:20+03:00), which publishes the chain result only after all seal checks complete. Tracked as a bug, not a security advisory. - [EGo: untrusted host can switch any enclave into a Marble and replace signed config files, env, and argv](https://uc.surf/findings/security/go/edgelesssys__ego/2026-10-13-host-forced-marble-mode-config-replacement/): edgelesssys/ego · Go · Medium · Published · reported 2026-10-13. An EGo app that was not meant to run under MarbleRun could still be started in Marble mode by a malicious host. With a fake MarbleRun Coordinator, the host could then supply a crafted configuration and take integrity control of the app's embedded files, environment variables, and arguments. The attestation evidence did not show whether the app ran standalone or as a Marble, so a remote verifier could not detect this. Fixed in 1.10.1, where Marble mode is an opt-in in enclave.json and changes the enclave's UniqueID. - [libpskc signature verification accepts embedded ds:KeyValue keys on xmlsec < 1.3.3](https://uc.surf/findings/security/c/oath-toolkit__oath-toolkit/2026-10-11-libpskc-keyvalue-signature-bypass/): oath-toolkit/oath-toolkit · C · Medium · Published · reported 2026-10-11. pskc_verify_x509crt leaves xmlsec keyInfoReadCtx.enabledKeyData unrestricted; on xmlsec < 1.3.3 a PSKC signed with an embedded attacker ds:KeyValue reports valid_signature=1 against an unrelated trusted certificate, enabling forged OTP seed provisioning. - [Missing handling for SVG animateTransform, animateMotion, and animateColor](https://uc.surf/findings/security/rust/rust-ammonia__ammonia/2026-10-05-animatetransform-animation-xss-bypass/): rust-ammonia/ammonia · Rust · Medium · Published · reported 2026-10-10. The neutralization ammonia applied to the SVG animate and set elements was not applied to animateTransform, animateMotion and animateColor, so these could set an href to a javascript: URL and lead to stored XSS in applications that allow those tags. Fixed in 4.1.7 and 4.2.3. Reported by email to the maintainer. - [Non-deterministic result from role policies due to cache collision](https://uc.surf/findings/bugs/go/cerbos__cerbos/2026-10-09-role-policy-cache-collision/): cerbos/cerbos · Go · Low · Fixed · reported 2026-10-09. Public issue opened by the maintainer and labelled as a bug (kind/bug), crediting ucsurf for the report. When a role policy had two resource entries that both matched the requested kind, their first rules got the same cache key, so the second entry's condition reused the first entry's cached result and was never evaluated. Fixed by the maintainers in PR #3410 (https://github.com/cerbos/cerbos/pull/3410, merged 2026-10-12T10:48:51+03:00); the issue was closed on 2026-10-14T10:47:47+03:00. Tracked as a bug, not a security advisory. - [bzip2-cli (bzip2 -d): decompressed output world-readable while written; setuid/setgid bits copied onto invoker-owned file](https://uc.surf/findings/security/rust/trifectatechfoundation__libbzip2-rs/2026-10-06-bzip2-cli-decompress-output-permissions/): trifectatechfoundation/libbzip2-rs · Rust · Medium · Published · reported 2026-10-09. uncompress() creates output without a restrictive mode and only chmods at the end (TOCTOU read window); set_permissions_rust copies S_ISUID/S_ISGID without fchown, yielding an invoker-owned setuid file. Regressions vs C bzip2. - [cron-parser: H/ (and H(min-max)/) field expansion is still unbounded at parse time (heap-exhaustion DoS), incomplete fix of GHSA-r9w7-75q6-r884](https://uc.surf/findings/security/javascript/harrisiirak__cron-parser/2026-10-06-hashed-step-expansion-heap-dos/): harrisiirak/cron-parser · JavaScript · Medium · Published · reported 2026-10-09. The 256-value cap added in 5.10.1 for GHSA-r9w7-75q6-r884 runs after hashed fields are expanded, so it does not cover H/ or H(min-max)/. Each such token expands to every value in its range before the cap applies, and a long expression of repeated tokens passed to CronExpressionParser.parse() can exhaust the Node.js heap and stop the process. Fixed in 5.10.2. - [preact-render-to-string: non-string attribute values skip HTML entity encoding (SSR XSS)](https://uc.surf/findings/security/javascript/preactjs__preact-render-to-string/2026-10-06-nonstring-attr-encode-bypass/): preactjs/preact-render-to-string · JavaScript · Low · Published · reported 2026-10-08. In the default renderToString and renderToStaticMarkup path, only string attribute values were HTML-entity encoded. Arrays and objects with a custom toString were converted to text and written into the attribute without encoding, so quote characters in them could break out of the attribute and inject markup during server-side rendering. This is distinct from the Preact core issue GHSA-36hm-qxxp-pg3m. Fixed in 6.8.1. - [svg-sanitizer: UTF-16 input bypasses DOCTYPE stripping (incomplete fix of GHSA-9rjx-3jch-6vjf)](https://uc.surf/findings/security/php/darylldoyle__svg-sanitizer/2026-10-04-utf16-bom-bypass-removedoctype/): darylldoyle/svg-sanitizer · PHP · Medium · Published · reported 2026-10-08. The fixes for GHSA-9rjx-3jch-6vjf (stored XSS through DTD entities) and GHSA-v383-3rw5-q8rf (crash through DTD attribute declarations) removed by searching the raw input bytes. libxml detects UTF-16 input from its byte order mark, and in UTF-16 the DOCTYPE bytes are interleaved with zero bytes, so the search found nothing and the internal DTD reached the parser. This brought back the DTD-based bypasses those fixes were meant to stop. Fixed in 1.1.0. - [templ: javascript: URLs not sanitised in other URL attributes and mixed-case attribute names](https://uc.surf/findings/security/go/a-h__templ/2026-10-06-url-attribute-sanitization-gap/): a-h/templ · Go · Medium · Published · reported 2026-10-07. templ sanitised string URLs only in a few attributes (a/link href, form action, object data) and only when the attribute name was lowercase. String values in other URL-bearing attributes, or in uppercase or mixed-case attribute names, were written without URL sanitisation, so javascript: URLs could pass through and run script. Fixed in v0.3.1100. - [templ: {{ }} interpolation in a script template literal not escaped for ${](https://uc.surf/findings/security/go/a-h__templ/2026-10-06-script-template-literal-interpolation-xss/): a-h/templ · Go · Low · Published · reported 2026-10-07. Inside a