Nous Research Releases Official Hermes Security Plugin: Parallel Subagent Reviews, Optional Test Confirmation, and SARIF Export

Core Background & Industry Pain Points

Coding agents can already edit a local repository, but traditional static scans often dump a pile of alerts without saying whether an attacker can actually reach the code. Nous Research's official account announced at 18:47 UTC on 2026-11-08 (02:47 SGT on November 9) that the official plugin hermes-security is now available. It was written by Hermes Agent engineer Mark S. (@mark_s; public bio: Building Hermes Agent @NousResearch, CISSP). The plugin reviews a local repository for security bugs. It requires Hermes 0.21.5 or later and Python 3.11 or later, and it is licensed under Apache-2.0.

Architecture Highlights & Internals

According to the official docs, the plugin first builds a threat model: what an attacker could reach and which inputs are trusted. It then splits the code by folder into chunks of about 4,000 lines. Up to 10 reviewer subagents read in parallel, and one more reviewer looks across the whole scope for entry points and trust boundaries. If production code fits in 90 chunks (about 360,000 lines), all of it is read; larger repositories are ranked by risky behavior such as routes, request input, shell commands, SQL, deserialization, auth and webhooks. There are standard, deep and diff modes. A deep scan defaults to three independent passes over the whole repository. Optional test confirmation writes small reproductions and runs them in a temporary copy of the repo. Testing a running app needs a per-scan grant (one origin, one hour, 200 requests) that agent tools cannot issue. The first scan downloads a pinned osv-scanner release of about 55 MB from GitHub and refuses it unless the SHA-256 matches.

Authoritative Benchmarks & Measured Scores

The vendor has not published recall, precision or finding counts for hermes-security against Semgrep, CodeQL or a public vulnerability corpus. The docs say the scan first checks lockfiles against published advisories with Google's osv-scanner and checks declared runtime versions against endoflife.date, but an advisory alone does not become a finding: the review must show that the code actually reaches the vulnerable behavior. Reports keep evidence, rejected candidates and coverage gaps, and they map to OWASP Top 10:2025 and ASVS 5.0.0; the vendor states this is not a certification. The review uses the model configured in Hermes, and the docs warn that the model can miss issues or misread code, so conclusions should be checked by a human.

Developer Hands-on Guide

Install with hermes plugins install hermes-security --enable, then restart Hermes. In chat you can ask to scan the repo, use /security, or run hermes security-review from the CLI. Hermes Desktop adds a Security page that starts turned off under Capabilities → Plugins. Reports export as Markdown, JSON, CSV and SARIF. Automatic continue via gateway injection is off by default and fires at most three times without progress. Test confirmation runs as your user in a temporary copy; the vendor says that is not a sandbox. The configured model provider receives the code under review, and scan data stays in the local plugin data folder.