Nous Research said on its X account on 8 November that Hermes Agent can now look for security bugs in code with a new official plugin. Hermes Agent is Nous Research’s AI agent, used in the terminal, the Hermes Desktop app and messaging services such as Telegram and Slack. Mark S. (@mark_s) at Nous built the plugin from first-hand experience of security reviews, Nous’s Teknium said.

Attack surface first, then subagents

Nous said the plugin first maps what an attacker could reach and which inputs the code trusts. It then splits the code across parallel subagents, separate agents that each read their own part. The plugin checks every candidate finding against the code and traces whether an attacker can actually exploit the bug.

The news site AICoder cites the plugin’s documentation: the code is split by folder into chunks of about 4,000 lines, and up to ten subagents read them at the same time. If production code fits in 90 chunks, or about 360,000 lines, all of it is read. In larger repositories, the code is ranked by risky behaviour, such as request input, shell commands, SQL and authentication. By default, the deep review goes through the whole repository three times, AICoder reported.

The plugin checks dependencies for known vulnerabilities with Google’s osv-scanner. An advisory alone does not make a finding, however. The review must show that the code actually uses the vulnerable function.

Tests run with the user’s permissions

With optional test confirmation, Hermes writes a small test that reproduces each finding and runs it in a temporary copy of the repository. The vendor notes that the copy is not a sandbox, AICoder reported: tests run with the user’s permissions. Testing a running app needs a per-scan grant, which the agent’s tools cannot issue themselves. The grant limits testing to one origin, one hour and 200 requests.

The code under review is sent to the model provider configured in Hermes. Scan data stays on the machine in the plugin’s own folder.

No comparative figures on accuracy

Reports show the evidence, the areas left unreviewed and rejected candidate findings. Users can save them in Markdown, JSON, CSV and SARIF formats. SARIF is a results format that other code analysis tools understand.

Nous has not published figures on the share of bugs the plugin finds or on how accurate its findings are, AICoder said. Nor is there a comparison with tools such as Semgrep or CodeQL. The documentation says a human must check the conclusions, because the model can miss bugs or misread code. Reports use the OWASP Top 10 and ASVS classifications, but the vendor says this is not certification.

The plugin is in Nous’s official plugin catalogue and can be used in chat, from the command line or from the Security page in Hermes Desktop. It requires Hermes version 0.21.5 or later and Python 3.11 or later, AICoder reported, and the first scan downloads osv-scanner, which is about 55 megabytes. Installation instructions are in Nous’s plugin repository on GitHub.