Chainguard's new Harness turns Claude Mythos 5.1 loose on live builds — only reports the bugs it can reproduce

Supply-chain security company Chainguard announced on Wednesday Chainguard Harness, a new service that unleashes frontier AI models — starting with Anthropic’s Claude Mythos 5.1 — on a live, isolated build of a customer’s application, reproduces every flaw until it proves itself, and delivers each finding with a ready-to-merge fix. The pitch: frontier-model vulnerability hunting for teams that don’t have frontier-model access.

Attack the live build, not the code

Static scanners read code; Harness plays attacker. Teams connect a repository through Chainguard’s Guardener GitHub app, and Harness compiles the application exactly as it ships, boots it inside an isolated microVM — no network egress, no embedded credentials, full audit logging — and lets the model probe the running instance in context. Nothing gets reported until the finding reproduces in a fresh, isolated environment on every replay, a design meant to kill the false-positive backlog that eats engineering hours. A Chainguard engineer signs off on every report and fix before a customer ever sees it.

No model access required

That last part is the business model as much as the safety story. Frontier models that perform vulnerability discovery across first-party and open-source code have, until now, been reachable mostly by a select group of organizations doing defensive work for critical infrastructure, Chainguard says. Harness hands that capability to everyone else — no frontier-model access needed on the customer side. Chainguard is selling it through the Claude Marketplace by Anthropic, so enterprises spend a slice of their existing Anthropic commitment and Anthropic handles invoicing, consolidating AI spend. A beta waitlist opened alongside the launch, and at launch Chainguard engineers operate Harness on customers’ behalf.

Why now: scanners drown teams in noise

The release reads like an indictment of the current toolchain. Traditional scanners analyze static code, flag issues that may never be reachable, and miss the logic flaws, injection paths, and resource-exhaustion bugs that only surface when an application is actually running. Engineers end up spending hours and tokens triaging unverified alerts and writing their own patches instead of shipping. Harness inverts the workflow: find only what’s exploitable, prove it, ship the fix. “Every time we’ve pointed Harness at code, it has found serious, exploitable vulnerabilities, including in software that’s been scanned many times over and code that runs critical infrastructure,” said CEO and co-founder Dan Lorenc.

Vulnerabilities Harness finds in open-source code go upstream to maintainers through Athena, Chainguard’s coordinated disclosure process — a neat reinforcement of its position as a trusted open-source supplier, since Chainguard can then ship patched builds through its own repository.

Assemble London, and the Claude Mythos 5.1 signal

The announcement landed at Chainguard’s Assemble London conference, where the company also unveiled Chainguard Workspaces — hardened, credential-free cloud environments for developers and their agents — and Chainguard Checks, which runs continuous-integration tests on code before a pull request is opened, plus Go and .NET support for Chainguard Libraries. The choice of model is its own signal: the press material names Anthropic’s Claude Mythos 5.1 as the frontier model doing the hunting, and Anthropic’s head of commercial cyber, Daniela Hurtado, is quoted endorsing the deployment model — isolation, reproduction, and human review around a model that can surface vulnerabilities “that have gone unnoticed for years.”

What we don’t know yet

Pricing is undisclosed, and a beta waitlist means limited availability for now. There’s no independent validation yet of hit rates or false-positive numbers outside Chainguard’s own claims, and the launch announcement doesn’t detail how Harness scales to sprawling monorepos or legacy stacks that won’t compile cleanly in a sandbox. Still, the direction is hard to argue with: verification as a product — sell the proof, not the alert. If Harness delivers on the reproducibility claim, it stops competing with scanners and starts competing with pentest firms.