- Chainguard says Harness uses Claude Mythos 5.1 to probe a running copy of an app, then reproduces each flaw before delivering a reviewed fix.
- The point for leaders: the scarce resource is no longer finding flaws but proving them and shipping fixes. Claims so far come from Chainguard and Anthropic.
- Before buying, ask for reproduction steps on a pilot, find out what repository access is granted, and check whether engineers can ship the fixes.
A list of maybes is not a finding
Most security teams do not lack alerts. They lack certainty about which alerts are real. Chainguard's new product, Harness, is built around that gap.
Chainguard announced Harness on October 14, according to Help Net Security. It uses frontier AI models, such as Anthropic's Claude Mythos 5.1, to look for exploitable flaws in an organization's own code. It is sold through the Claude Marketplace.
The idea worth noticing is this. When AI makes finding flaws cheap, proof becomes the scarce thing. A flaw nobody can reproduce is an opinion. A flaw that reproduces is a work order.
How Harness works
Harness connects to a team's repository through the Guardener GitHub app. It builds the application exactly as it ships. It then boots that build inside a microVM, which is a small, locked-down virtual machine.
A frontier model probes the live application, testing it in context the way an attacker would. Chainguard says this finds problems that static code review misses. Its examples are logic flaws, injection paths and resource exhaustion. Injection means outside input gets treated as a command. Resource exhaustion means crafted input drains an app's memory or capacity.
Then comes the step that matters most. Chainguard re-runs each finding in a new, separate environment. A finding that does not repeat on every attempt does not move forward. Only proven findings reach the customer, with reproduction steps and a fix written for that codebase. A Chainguard engineer reviews every fix and report.
Why the safeguards are part of the product
Chainguard lists the limits it puts on each run. The sandbox cannot reach out to the internet, holds no credentials, and keeps a complete audit trail. Anthropic's Hurtado described the design as surrounding Mythos 5.1 with "isolation, reproduction, and human review."
This matters because the tool is, in effect, a model that behaves like an attacker, aimed at your own software. The value comes from the controls around it. Chainguard also says flaws found in open source code go upstream to maintainers through its disclosure process, called Athena.
What is not yet proven
These are the vendor's claims. Chainguard CEO Dan Lorenc says that each time Harness was pointed at code, it found exploitable vulnerabilities. He includes software that had been scanned many times. The announcement gives no figures on how many flaws, in which products, or how often.
Harness is also not self-serve at launch. Chainguard engineers operate it. They run it against the open source artifacts Chainguard distributes, and against a customer's source code, delivering a reviewed report. That makes it closer to a managed service than a scanner you switch on.
The human cost it targets
Chainguard's pitch points to a familiar drain. Engineers spend hours and AI tokens sorting through alerts and drafting their own patches, and shipping waits. That is the person behind the backlog. Proof changes their day: they open a ticket with steps that work and a fix that fits their code.
The catch is that a proven backlog is still a backlog. Verified findings are harder to ignore. If your team cannot ship fixes faster, better evidence will not help.
Questions to put to your team
First, ask what repository access the Guardener app needs, and who at the vendor can see your source. Second, ask for reproduction steps on a pilot and have your own engineers replay them. Third, ask who tests each fix before it merges.
Fourth, ask how disclosure timing works when a flaw sits in open source you depend on. The announcement does not give a timeline.
Finally, note how it is bought. Customers can pay through their Anthropic commitment, which removes a procurement step. That is convenient. It also means a security decision can pass through the AI budget, so name who owns the review of what it delivers.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Help Net Security.
