Chainguard Harness: How AI Finds and Fixes Bugs in Your Code

Chainguard Harness is a new service that puts frontier AI to work finding exploitable bugs in application code, then proves each one is real before anyone sees it. Announced October 14, Harness uses Anthropic's Claude Mythos 5.1, operated by Chainguard's own staff, to probe a customer's running application inside an isolated sandbox. Every finding is replayed until it reproduces reliably, comes with a tested fix, and is approved by a named Chainguard engineer before delivery.

• Model: Anthropic's Claude Mythos 5.1, run by Chainguard staff, so customers do not need their own model access.
• Isolation: apps are rebuilt from source and tested in a microVM with no network access and no credentials.
• Verification: each exploit must reproduce every time in a fresh microVM, and each fix must make the exploit fail.
• Availability: a beta waitlist is open now, and Harness is sold through the Claude Marketplace.

How Does Chainguard Harness Work?

Customers connect their repositories through a read-only GitHub App or upload a source archive. Chainguard then rebuilds the application from source, the way it would ship to production, and boots it inside an isolated microVM.

The process runs in two phases, according to Chainguard's announcement. A separate model handles setup, and then the microVM is suspended. For the hunting phase, it resumes with no network egress and no credentials, and the model gets only a shell and Chainguard-built packages. The agent loop itself runs outside the sandbox, so every tool call is logged, and the code under test is treated as untrusted data.

During the hunt, the model probes the live application for runtime and logic flaws, injection bugs and resource exhaustion, drawing on the source code, git history, dependency source and build configuration.

Why Does Exploit Replay Matter?

Anyone who has run a traditional scanner knows the problem: long reports full of findings that turn out not to matter. Harness takes the opposite approach. As Help Net Security reports, only confirmed findings are delivered, each with reproduction steps and a fix tailored to the customer's codebase.

The final report includes an executive summary, a severity and impact rating for each issue, steps to confirm it, and the fix, while describing issues rather than handing over a working attack. A Chainguard engineer reviews and approves every fix and report. That combination of AI scale and human sign-off is what makes the results something a security team can act on immediately.

What Happens When Harness Finds an Open-Source Bug?

When a flaw lives in an open-source dependency rather than the customer's code, Chainguard discloses it to the upstream maintainers through a coordinated process, then ships the patched build through its own repository once it is fixed. Chainguard has already shared an example: its harness independently flagged a vLLM issue where one client's output restriction could carry over to another client's request. Chainguard reported the security impact on September 23, the fix shipped in vLLM v0.31.0 on October 2, and an advisory followed on October 6.

This builds on a theme we have been tracking all year, as AI-assisted discovery programs like Project Glasswing help maintainers find and patch bugs faster.

How Can Teams Get Chainguard Harness?

Harness is in beta with access initially limited to a small group of organizations, and Chainguard has opened a waitlist. Because it is available through the Claude Marketplace, organizations can apply part of an existing Anthropic commitment toward it. Chainguard has not published pricing. Chainguard CEO Dan Lorenc told Help Net Security that Harness has found serious, exploitable vulnerabilities every time it has been pointed at code, which is a strong signal of how much defenders stand to gain from verified, AI-driven testing.

For more on tools that help defenders stay ahead, see our AI security coverage.

Sources: Chainguard, Introducing Chainguard Harness — October 14, 2026; Help Net Security, Chainguard Harness — October 14, 2026; Chainguard, The first disclosure from our Mythos harness — October 13, 2026.