Hermes Security adds AI code reviews, with testing limits
Hermes Security gives Hermes Agent a dedicated workflow for reviewing a repository, tracing possible attack paths and keeping the evidence behind its conclusions. Nous Research announced the official plugin on November 8, 2026. Developers can start a review from chat, the command line or a Security page in Hermes Desktop.
The November 8 announcement, reproduced on Nous Research’s website, describes an agent that maps trusted inputs and reachable code before dividing the review among parallel agents. It then checks potential findings against the repository. We reviewed the published documentation, not a running installation, and have not measured how reliably it finds bugs.
What the review produces
According to Nous, optional test confirmation creates small reproductions in a temporary repository copy. Reports preserve supporting evidence as well as candidates the review ruled out, and can export to SARIF for other security tools.
That record matters when deciding whether a finding deserves an engineer’s time. A plausible warning, a reachable vulnerable path and a successful reproduction answer different questions. For example, a risky function may exist in a repository without being reachable through the deployed application’s inputs. A useful review needs to connect those steps.
Before accepting a result, ask which input an attacker controls, where it enters the application and what code permits the claimed outcome. Keep the original evidence with the issue so another engineer can check it after a fix. These are our suggested review questions, not results from testing Hermes.
A separate plugin, with separate updates
The official plugin repository explains that integrations ship independently of the Hermes core. Catalog entries point to reviewed commits, and updates move installations to a new approved pin. The repository also describes a partial download that retrieves the chosen plugin rather than every integration.
Record both the Hermes version and the plugin revision when comparing scans. Updating the host application alone is not a reliable record of which security reviewer ran. Keep that information alongside the repository commit under review.
Our Hermes v0.23 guide covers browser access, conversation rewind and spending controls. Those changes belong to the broader agent. This announcement concerns a distinct review workflow.
Check permissions before enabling tests
The marketplace documentation requires Hermes 0.21.5 or later and Python 3.11 or later. It lists standard, deep and diff reviews. Its listing predates the announcement, so November 8 should not be read as the first day the code existed.
The most important warning concerns execution. Optional reproduction tests run with your user permissions. The temporary copy and removal of environment variables resembling credentials do not provide a sandbox. Testing a running application requires permission granted by the user for one origin, one hour and up to 200 requests.
The configured model provider receives code under review. Dependency checks send package metadata to external advisory services. Reports stay in the active Hermes profile, but local storage does not make the analysis private to that machine.
A practical first trial should use a repository you are authorized to share with that provider and an isolated environment containing no production credentials. Decide separately whether to allow code execution. An installation that is convenient for ordinary coding may be an inappropriate place to run unfamiliar test programs.
Read coverage before counting findings
The documentation says suggested patches are not automatically applied. It also warns that missing coverage and model mistakes limit the conclusions. Security standard mappings are not certification.
Start with one reviewable change and compare the report with an engineer’s assessment. Check whether the report identifies the intended scope, preserves enough evidence to reproduce its reasoning and clearly marks unfinished work. A clean summary is less useful than an honest account of what was actually examined. The value to look for is a shorter path from a warning to a defensible engineering decision.
Featured image: AI-generated editorial illustration by ByteForward. Hermes Agent wordmark: Nous Research, copyright 2025, from its MIT-licensed repository. Conceptual artwork, not a product screenshot.
