Chainguard Harness lets AI attack your app before hackers do
Chainguard uses Anthropic's Claude Mythos 5.1 to hack your live app in an isolated sandbox, but a human engineer approves every fix.
Tools · Source: SiliconANGLE
What happened
Chainguard just launched Harness. It is a vulnerability-hunting service that uses frontier artificial intelligence models like Anthropic's Claude Mythos 5.1 to attack your code. Instead of reading static files, Harness compiles your application exactly as it would ship. It boots the build in an isolated microVM cut off from the outside network. The AI then probes the live, running instance exactly like a human attacker would.
The system is built to eliminate security noise. Harness only reports a vulnerability if it can successfully reproduce the exploit in a fresh environment on every replay. Once a flaw is proven, the system generates a custom fix for your specific codebase. A human Chainguard engineer reviews and signs off on every single report and fix before you see it. Flaws found in open-source projects are sent upstream to maintainers through a disclosure coalition called Athena.
You do not need direct access to frontier models to use this service. Harness is available directly through the Claude Marketplace. Enterprises can use their existing Anthropic spending commitments to buy it. The service is currently operating under a beta waitlist. Chainguard also announced new tools to secure AI coding agents, including credential-free cloud environments called Workspaces, and expanded its vetted open-source libraries to Go and .NET developers.
Key facts
- Mythos 5.1 — The Anthropic Claude frontier model powering Chainguard Harness.
- $892 million — Total venture capital raised by Chainguard since its founding in 2021.
- 5 — Number of languages now supported by Chainguard Libraries, adding Go and .NET.
Why it matters
Traditional static scanners are drowning engineering teams in false positives. They flag unreachable code and completely miss runtime logic flaws, injection paths, and resource exhaustion bugs. Harness changes the paradigm from static analysis to live, agentic penetration testing. You stop wasting tokens and engineering hours triaging unverified alerts. You only look at proven vulnerabilities that come with a ready-to-merge, human-reviewed patch. This keeps your team focused on shipping features instead of chasing ghosts.
Frontier AI models are incredible at chaining exploits, but access has been restricted to a small group of defensive security teams. Chainguard is democratizing this capability. By wrapping the AI in a secure sandbox and handling the deployment, they are making elite red-teaming accessible to standard enterprise teams. This forces attackers to work harder. Companies can now patch zero-day logic flaws before they even ship to production. As AI coding agents write more code, automated AI defensive tools like this become mandatory.
For builders
Stop chasing static scanner ghosts
Static scanners flag issues that are never reachable in production. Harness only reports flaws it can reproduce on a live build, saving your engineers hours of useless triage.
Consolidate your enterprise AI spend
Harness is listed on the Claude Marketplace. If you have an existing Anthropic spending commitment, you can use it to procure this service without going through a separate vendor approval process.
Secure your AI coding agents
Chainguard also launched Workspaces and Checks. These give your coding agents hardened, credential-free environments and run continuous integration tests before pull requests are opened.
My take
AI is great at hunting, but I do not trust it to make the final call on my codebase. Chainguard gets this exactly right. By forcing the AI to prove the exploit and requiring a human engineer to sign off on the fix, they deliver actual security instead of just more AI noise.
Original reporting: SiliconANGLE. This is my rewrite and opinion.