Chainguard has launched Chainguard Harness, a cyber harness that applies frontier AI models such as Anthropic's Claude Mythos 5.1 to live application testing. The service builds an organization's code in an isolated microVM, lets the model probe the running instance for exploitable flaws, and replays each finding in a fresh environment until it reproduces consistently. Only then does it deliver a fix tailored to the codebase, with human review by a Chainguard engineer.
According to the announcement, the approach targets runtime issues that static scanners miss, such as logic flaws, injection paths, and resource exhaustion. Teams do not need their own frontier-model access; Harness is available through Anthropic's Claude Marketplace, so enterprises can apply existing Anthropic commitments and consolidate AI spend. Chainguard also routes open source findings upstream through its Athena disclosure process.
Chainguard CEO Dan Lorenc is quoted saying every run against code has found serious exploitable vulnerabilities, including in heavily scanned software and critical infrastructure. Anthropic's Daniela Hurtado emphasizes isolation, reproduction, and human review as safeguards around the model's capability. The source is a single vendor announcement, so independent validation of these claims is not included.