Why it mattersA team that cannot access the restricted Claude Mythos model can now get its findings as a managed service, and the vendor says every bug is reproduced in a clean build before it reaches a customer.
Most companies with revenue above one billion dollars know about critical bugs in their software and ship anyway. Chainguard just launched a product that uses Anthropic's most restricted model to find them, and the first bug it disclosed was quiet in a way scanners miss.
Chainguard Harness is a managed vulnerability-hunting service that builds the customer's application from source, runs it in an isolated microVM, and lets Claude Mythos 5.1 attack the live instance. It is one of four products the company announced, alongside Workspaces, Checks and expanded Libraries coverage. Harness is also available through the Claude Marketplace, so existing Anthropic spending commitments can pay for it.
Why access to Mythos matters
Anthropic restricts Claude Mythos 5.1 to companies participating in its Project Glasswing. Chainguard tells The New Stack it first built Harness for its own use of Mythos inside Glasswing, and is now opening the harness up to customers who could not otherwise touch the model. The company says Harness runs on Mythos today but is meant to be model-agnostic.
Every finding goes through a reproduction step before it reaches a customer. The service proves a vulnerability is exploitable by running the exploit against a clean build, and a proposed patch has to succeed against the original code and fail against the patched version before the fix moves on. A Chainguard engineer then approves every finding and every fix.
The vLLM bug
The first public disclosure from Harness is a flaw in vLLM, the open-source engine many teams use to serve large language models. In versions before v0.31.0, one client could force the response returned to another client sharing the same engine, and the tampered reply came back looking like a normal, successful completion with nothing in the logs. The bug is fixed in v0.31.0 and tracked under advisory GHSA-6cxc-2vcg-w5qc. Chainguard says further findings, including some in critical infrastructure, are under coordinated disclosure.
The survey behind the launch
Chainguard commissioned Wakefield Research to survey 1,300 security leaders across nine markets, in a report titled The AI Exploit Gap. More than half of the respondents said their organisations knowingly leave high or critical vulnerabilities unpatched at least sometimes. At companies with one billion dollars or more in revenue, that figure is 66 percent. At companies below 250 million dollars in revenue, it is 27 percent. Quincy Castro, Chainguard's chief information security officer, said in a statement that AI can find and chain flaws in minutes while fixing them still moves at human speed.
The three other products shipped the same day work on the agent side. Workspaces runs each developer and coding agent in a hardware-isolated microVM, keeps credentials outside the agent's environment, and signs the activity log. Checks lets coding agents run CI tests on signed, short-lived credentials in seconds, from warm snapshots that refresh when the lockfile changes, so an agent can run tests, inspect failures and revise code before it opens a pull request. Libraries adds Go and .NET package endpoints, with CVE remediation for JavaScript and Go.
A team adopting an AI bug finder has had to pick between vendor marketing and running the model themselves. A managed service that gates every bug through a reproduction step, and ships patches through a signed package endpoint, moves the review burden off the team and still leaves a signed audit trail.
Source
Introducing Chainguard Harness, Chainguard. Reported by Amanda Caswell in 66% of billion-dollar companies leave critical bugs unpatched, The New Stack.
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 using them to release software. Short, and only when there is something worth reading.
