Chainguard puts Claude Mythos on troubleshooting, securing coding agents and expanding libraries

Chainguard Inc., a provider of secure software supply chain solutions, today introduced a vulnerability hunting service that leverages novel artificial intelligence models such as Anthropic PBC’s Claude Mythos 5.1, as well as two products for securing AI coding agents and an expansion of the Chainguard libraries.

All four attended the company’s Assemble conference in London today.

Vulnerability hunting is the responsibility of Chainguard Harness, which only reports the bugs that it can reproduce from a live build of a customer application. Chainguard Workspaces provides developers and their agents with a hardened, credential-free cloud environment.

Because this work still needs to be tested, Chainguard Workspaces has combined with Chainguard Checks, which runs continuous integration tests on the code before anyone opens a pull request. Go and .NET developers are the new audience for Chainguard Libraries, the company’s service for verified open source dependencies.

Harness is aimed at organizations that want to look for vulnerabilities in boundary models without direct access to the models themselves. According to Chainguard, these models have largely gone to a small set of groups doing defensive security work, and static scanners miss the logic errors that only surface once an application is running.

Customers connect a repository via Chainguard’s Guardener GitHub app. From there, Harness compiles the application as it would be delivered, and the build starts in an isolated virtual machine cut off from the external network. A Frontier model then examines this running copy the way an attacker would. According to the company, nothing will be reported until it is reproduced in a new environment each time it is played. A Chainguard engineer signs every report and correction before a customer sees it.

“Each time we have pointed out code to Harness, serious, exploitable vulnerabilities have been found, including in multi-scanned software and code that runs critical infrastructure,” said co-founder and CEO Dan Lorenc.

Currently, Chainguard’s own employees operate harnesses for customers. Scans of the open source artifacts the company distributes result in patched builds delivered via the Chainguard Repository.

When Harness is run using a customer’s own code, the customer receives a description that includes severity, replication steps, and a recommended solution. Defects discovered in open source projects are passed on to upstream maintainers via Athena, a disclosure coalition that Chainguard founded in June.

Anthropic also sells harnesses. The service is listed on the Claude Marketplace, which opened in March, and companies can use a portion of an existing Anthropic spending commitment toward it. Daniela Hurtado, head of commercial cyber at Anthropic, said how a model like Mythos 5.1 is deployed “is as important as the capability itself.”

Harness looks for errors in the software that a company already ships. The two agent products intervene earlier, at the point where the coding agents are still writing the code. Chainguard Workspaces and Chainguard Checks are designed for teams whose agents perform long migrations with the same permissions as the developer who started them, within reach of cloud keys and production systems. Approval prompts rely on the agent following the rules, and a developer juggling a dozen agents can end up clicking “allow” all day.

Each workspace is a remote virtual machine built entirely from Chainguard OS packages. By syncing with the developer’s local working tree, the editor and Git workflow stay on the laptop while execution occurs remotely.

Real credentials remain outside the machine. Tools that require one on disk receive an encrypted, short-lived replacement, and teams can enable outbound policies that block any connections they haven’t approved. Lorenc described the result as an isolated machine “with nothing in it worth stealing.”

The checks target the queue that arises when agents push code much more frequently than continuous integration systems are designed for. Each run starts with a warm snapshot with dependencies already installed. An agent working in a workspace will have its session paused during testing and then resume, with all error logs fed directly into its context.

Passing runs produce a signed attestation tied to the environment, test result, and commit. Merge gates can verify this attestation and merge the code without having to re-run the suite.

James Governor, co-founder of analyst firm RedMonk LLC, said existing continuous integration platforms are “struggling with the massive increase in burden” that agent development has brought. The Australian payment infrastructure provider Cuscal Ltd. is one of the first design partners for both products.

The library update expands the service to five languages. Java was the only supported language library when Chainguard launched the service in March 2025, and Python and JavaScript have since been added. Teams building on .NET can now swap NuGet for a Chainguard managed endpoint.

Go and .NET both have configurable blocking policies and malware scanning that also catch greyware. Chainguard applies the term to packages that work as advertised while performing malicious actions, such as collecting credentials.

Chainguard makes the Go libraries available to teams developing cloud and infrastructure software in the language. Older Go packages also receive backported upstream fixes when a known critical or high-severity vulnerability is found, sparing their assigned teams a forced major upgrade. JavaScript libraries now provide the same workaround.

Java and Python receive zero-day protection based on insights pooled by Athena members. Chainguard says patched artifacts for bugs found by Frontier AI models will be ready the day they are released.

Workspaces and checks are available in beta testing starting today. Harness is open for its own beta via a waitlist, and early access has begun for the Go and .NET libraries.

Chainguard is a venture capital-backed startup that has raised approximately $892 million since its inception in 2021. The most recent $280 million came from General Catalyst’s Customer Value Fund last October.

Image: SiliconANGLE/GPT Image 2.5