Chainguard has introduced a set of tools aimed at the security problems created when AI agents write and test software. At its Assemble London event on October 14, the company announced Harness, a service that uses Anthropic's Claude Mythos 5.1 to look for exploitable vulnerabilities in a running application. It also announced isolated development environments and checks for code written by people or agents. The company's announcement says Harness, Workspaces and Checks are in beta, with waitlists open.
Testing software as it runs
Chainguard says Harness builds an application from source, starts it inside a microVM without credentials or network egress, and lets the model probe the running system for issues such as logic flaws, injection and resource exhaustion. Candidate exploits are replayed in a fresh environment, and suggested fixes are tested against the same behavior. A Chainguard engineer reviews the resulting report before delivery. These are the company's stated controls and workflow; independent data on the service's detection rate or false-positive rate has not yet been published with the launch.
There is one public example of the research approach behind the service. Chainguard described a vLLM issue in which one client's request could influence another client's output. The company says it independently found the problem, notified maintainers of its security impact in September and saw a fix ship in vLLM 0.31.0 on October 2. That example supports the claim that its researchers have used a Mythos-based harness in coordinated disclosure, but it does not establish how the new commercial service will perform across other applications.
Containing agents that write code
Chainguard Workspaces gives developers and agents separate microVM-based environments. According to the company, agents can run commands and install packages there without direct access to the developer's laptop or its long-lived credentials. Short-lived credentials can be supplied through a host-side proxy, while optional egress rules restrict network destinations. Chainguard Checks runs pre-commit and CI-style checks in similarly isolated environments. Both offerings are in beta, so their operational claims should be evaluated against actual deployment requirements rather than treated as universal protections.
The wider release also expands Chainguard Libraries into Go and .NET in early access. Together, these announcements address the full path from dependencies to development sessions to vulnerability testing. For security teams considering the products, the useful questions are specific: which credentials can an agent reach, what network access does it retain, how are findings reproduced, and what evidence is provided before a fix is trusted?