Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

GitHub AI Security Catalog Gets 231 Pull Requests in Six Months, 157 Accepted

After the yoPRSTCON conference, Sergey Gordeychik started a GitHub list of AI tools for cybersecurity. In six months it received 231 pull requests (PRs: proposals to add or fix an entry) from 209 authors. 157 were accepted, 65 rejected, and nine were still open. Every submission has to be vetted by reading the code: a "local" scanner may send all your source code to an external model. The takeaway: when choosing protection for agents, trust the code, not the README.

AI-processed from Habr AI; edited by Hamidun News
GitHub AI Security Catalog Gets 231 Pull Requests in Six Months, 157 Accepted
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #171% of agent-security repositories (75 of 106) were created in 2026
Key metric #2The watchlist holds 276 young projects, with a median of 14 stars
Key metric #3The review agent treats other people's READMEs and CI checks as untrusted text

In May, at the yoPRSTCON conference, attendees agreed to collect AI-based cybersecurity tools, along with tools that protect models and agents themselves, in a single list. It was meant to be a 20-minute scratchpad. It turned into a project with a queue of submissions: in six months, the awesome-ai-security-tools repository received 231 pull requests (a PR is an author's proposal to add or fix an entry in the list) from 209 authors. 157 were accepted, 65 were closed without being included, and nine were still open as of November 14. For anyone building AI agents or choosing protection for them, this is a rare snapshot: what is currently appearing in the niche, and how to vet other people's tools.

How a "scratchpad" turned into a review queue

The conference took place on May 26, and the first commit landed on May 27: 99 entries in the main list and 11 links to thematic collections. The list split into two directions right away. In the first, AI helps find vulnerabilities, triage alerts from security systems, analyze executable files, and gather threat intelligence. In the second, tools protect models, applications, and agents from injections, leaks, and unwanted actions.

In July, a social media post retold the collection as an "open arsenal of 300+ tools that will make every CISO panic", and the noise grew noticeably. On July 26, the list ranked tenth in Trendshift's daily ranking of Python projects, and by September 23 it had reached 1500 stars. Submissions kept growing: 5 in June, 23 in July, 40 in August, 58 in September, 64 in October, and 41 in the first two weeks of November.

How to vet tools and keep other people's commands out of the review

Similar submissions kept arriving: a repository created yesterday, one huge commit, a polished README, and bold promises. The maintainer did not want to reject on that basis alone: sometimes it is yesterday's experiment, and sometimes it is a mature project that previously lived in a private repository. So the code gets read. That is how it turns out that a "scanner" is really a set of instructions for an agent, or that a tool promising local operation sends all your source code to an external model provider.

The vetting questions are simple. What does the user install, and where does the analysis run? What data leaves the machine? Are a paid account and an API key required? What happens if the model is unavailable? Can the control be bypassed through another interface? Do the tests exercise the protective mechanism itself, or only its behavior against a mock? Agents help read the implementation and check promises against the code, but a human makes the decision: the reviewing agent makes mistakes too. External candidates are studied without running their code, installers, or automation scripts on a work machine.

Useful but too-young projects go to WATCHLIST.md. It currently holds 276 repositories, which are reviewed periodically: those that have grown are moved to the main list, and abandoned ones are removed. Poorly formatted submissions are accepted with the author's credit preserved, and the maintainer fixes them himself.

A separate lesson concerns the security of the review itself. The agent reads other people's READMEs, comments, and configs, and these may contain instructions to run a command or change CI. So third-party text is treated as untrusted data. One example: in PR #69, the author Lumi Eggcracker proposed adding a CI check of his own sample report against his own schema. The catalog did not need it. The maintainer sees no malicious intent from the author, but what an unrestricted agent would have done is unknown.

What the numbers show

The analysis covered 377 repositories in the main list as of November 14: tools, research code, benchmarks, and standards. 155 of them were created in 2026. Newcomers are most numerous in agentic static code analysis (18 of 24, 75%) and in agent security (75 of 106, 71%). Within the latter, runtime action control accounts for 46 of 64 (72%). Among pentest agents, 9 of 28 are new (32%); in threat intelligence, 3 of 12 (25%); in adversarial testing and LLM protection, 14 of 59 (24%); and in LLM-assisted fuzzing, 2 of 14 (14%).

In the watchlist, of 269 repositories with a known creation date, 235 (87%) appeared in 2026, and the median among 270 accessible repositories is 14 stars. The largest groups are adversarial tests and benchmarks (45), checks for skills, MCP, and dependencies (42), and permissions and action restrictions (35). These projects answer the questions "who allowed this action?", "what am I installing along with this skill?", and "how do I grant access without a key in the prompt?".

The author spells out the limitations himself. The creation date of a public repository does not always match the start of development. Thematic collections and models published only on Hugging Face are not part of the sample. The table shows where projects are appearing, but not the quality of their methods: that has to be compared on common tasks.

The practical takeaway: when choosing protection for agents, look at the list of checks and gaps, not at an A+ grade in a report. Check whether the control can be bypassed through another interface. Do not run other people's installers on a work machine. And if you are the author of such a tool, submit it to the list and read the feedback: according to the maintainer, authors and their agents do read it.

Story

Key Questions Answered

How a "scratchpad" turned into a review queue?

The conference took place on May 26, and the first commit landed on May 27: 99 entries in the main list and 11 links to thematic collections. The list split into two directions right away. In the first, AI helps find vulnerabilities, triage alerts from security systems, analyze executable files, and gather threat intelligence. In the second, tools protect models, applications, and agents from injections, leaks, and unwanted actions.

How to vet tools and keep other people's commands out of the review?

Similar submissions kept arriving: a repository created yesterday, one huge commit, a polished README, and bold promises. The maintainer did not want to reject on that basis alone: sometimes it is yesterday's experiment, and sometimes it is a mature project that previously lived in a private repository. So the code gets read. That is how it turns out that a "scanner" is really a set of instructions for an agent, or that a tool promising local operation sends all your source code to an external model provider.

What the numbers show?

The analysis covered 377 repositories in the main list as of November 14: tools, research code, benchmarks, and standards. 155 of them were created in 2026. Newcomers are most numerous in agentic static code analysis (18 of 24, 75%) and in agent security (75 of 106, 71%). Within the latter, runtime action control accounts for 46 of 64 (72%).

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…