Chainguard puts Claude Mythos on vulnerability hunting
Chainguard wants to rent out the frontier models that until now only a handful of defensive-security teams could touch — and to put a wall around the AI coding agents writing software in the first place.
At its Assemble London conference today, Chainguard launched Harness, a service that points frontier models such as Anthropic's Claude Mythos 5.1 at a customer's code the way an attacker would, and reports only the flaws it can actually reproduce. Customers connect a repository through Chainguard's Guardener app; Harness then compiles the application the way it would ship, boots that build in an isolated virtual machine cut off from the network, and lets the model probe the running copy. Nothing is reported unless it reproduces in a fresh environment on every replay, and a Chainguard engineer signs off on each finding before the customer sees it. Harness is also listed on the Claude Marketplace, the storefront Anthropic opened in March, which means enterprises can pay for it out of an existing Anthropic spending commitment — Chainguard's own staff run the scans for now, and a waitlist beta is open.
The pitch rests on a gap that static scanners have never closed: logic flaws that only surface once an application is running. Chainguard argues that frontier-model vulnerability hunting has mostly gone to a small set of well-resourced defensive groups, and that Harness is a way for everyone else to get the same reach without direct access to the models.
The other half of the launch targets agents rather than finished code. Workspaces gives coding agents a hardened remote machine built entirely from Chainguard OS packages, with real credentials kept off the box and tools handed short-lived encrypted stand-ins instead. Syncing with the developer's local working tree keeps the editor and git workflow on the laptop while execution happens remotely, and admins can block any network destination they have not approved. Paired with it, Checks runs continuous integration tests before anyone opens a pull request, starting from a warm snapshot with dependencies already installed and suspending the agent's session during the run so failure logs land back in its context. Passing runs produce a signed attestation tied to the environment, test result and commit, which a merge gate can verify without running the suite again. That is a direct answer to the queue problem agentic development creates — RedMonk co-founder James Governor said existing CI platforms are "struggling under the massive increase in load" that agents have brought. Australian payments infrastructure provider Cuscal is an early design partner on both products, which are in beta from today.
Chainguard also pushed Libraries, its service for vetted open-source dependencies, to five languages with Go and .NET early access joining Java, Python and JavaScript, adding configurable blocking policies and malware scanning that also catches packages which work as advertised while quietly harvesting credentials. The company has raised about $892 million since 2021, including $280 million from General Catalyst's Customer Value Fund a year ago — and it is now selling alongside Anthropic rather than against it.
What to watch: whether "reproduces on every replay" becomes the standard other AI bug hunters have to meet, and whether enterprises accept a frontier model reading their source at all.
Would you let a frontier model loose on your production build? Tell us in the comments.
