Why it mattersAn AI agent asked to hunt bugs will flood the report with plausible false positives unless the pipeline forces reproduction, and this project builds that constraint into the framework rather than into the prompt.
Zyrexnn's Cybermes, released as v3.3.0 on GitHub and updated on 5 September 2026, is an offensive-security framework designed to be driven by a coding agent rather than by a human at a shell. It ships a native Go MCP server, cybermes-mcp, that exposes more than ten security tools and context providers over JSON-RPC 2.0 to any editor that speaks the Model Context Protocol, including Claude Desktop, Cursor, Windsurf and Kilo Code. The repository is Apache 2.0, has taken 762 stars in the twenty days since it was created, and has been forked 133 times.
The proof-of-concept gate
The maintainer's own README argues that dropping a coding agent onto a target and asking for vulnerabilities is easy, and that keeping the false-positive rate low is the whole engineering problem. Cybermes tries to solve it in the framework rather than in the prompt.
Every finding has to clear a gate the project calls the Zero-False-Positive PoC Gate. Before anything reaches the findings/ directory the agent must produce a standalone, non-destructive Python script under pocs/poc_<name>.py, plus a raw HTTP trace, and both must reproduce the issue. A finding without a running PoC does not get logged, which turns the reporter into a check on the reasoner.
The Go binaries wrapped around subfinder, nuclei and sqlmap
The Go binaries do the boring work. smart_pipe streams high-signal endpoints, status codes, and leaked secrets out of noisy fuzz and crawl output, and archives the rest to disk so the model does not see it. secret_scan walks the raw recon directory for credentials that got captured. search_knowledge promises sub-50 millisecond offline lookup across 200 standard-operating-procedure playbooks, PayloadsAllTheThings, and HackTricks datasets, so a query for a technique like JWT algorithm confusion resolves without a web fetch. aggregate_reports compiles the confirmed findings into SUMMARY.md, a metadata.json metrics file, a standalone HTML dashboard, and a PDF named per target.
Under the top layer sit the familiar names. subfinder, httpx, katana, ffuf, nuclei and sqlmap are pre-wired, so the agent gets asked what to test and the harness supplies the toolchain. Every target creates its own directory under reports/<TARGET_SLUG>/, and the workspace isolation stops context from one engagement leaking into the next.
Two ideas worth stealing, even for teams already on Semgrep or Burp
The two design choices worth reading past the marketing copy for are the PoC gate and the stream filter. The gate forces the reasoner to earn the report row: the agent that hallucinated the SQL injection has to write the script that exploits it, and if the script does not work the finding is dropped. The stream filter tackles the other cost, which is that a raw nuclei run against a large target can burn tens of thousands of context tokens telling the model about 404s, and a filter that runs before the pipe saves both the bill and the reasoning quality.
Neither of those ideas is unique to Cybermes: the pipeline pattern of "recon streams, tools verify, agent narrates" is emerging across several open-source frameworks. What is new here is packaging it as an MCP server that any editor can call, licensing it under Apache 2.0, and shipping cross-platform installers for Windows, Linux, macOS and Docker with a doctor.py script that checks the environment before the first scan.
Two limits are worth stating plainly. This is offensive security tooling and the disclaimer at the bottom of the README says the framework is for authorised bug bounty, authorised penetration testing and internal red teaming. And Cybermes brings its own tools and playbooks, but a security team already committed to Semgrep, CodeQL, Burp Suite Pro or an internal SAST platform is not obviously replaced by it. The frame here is to give an agent the same tools a red teamer would run, on top of whatever platform already sits underneath.
Source
Source: GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually shipping with them. Short, and only when there is something worth reading.