66% of billion-dollar companies leave critical bugs unpatched — the part the AI bug-hunting hype leaves out

A new AI-powered security harness called Chainguard Harness has been unveiled by Chainguard, leveraging frontier models such as Anthropic's restricted Claude Mythos 5.1 to identify and rectify exploitable vulnerabilities within application code that traditional scanners often overlook. Developers can utilize Harness without direct access to Mythos, yet the specifics of which models are responsible for vulnerability detection remain undisclosed.

The company asserts that Harness has uncovered critical vulnerabilities within codebases that have already undergone rigorous security scans, including those employed in vital infrastructure. Information regarding the affected projects, the exact nature of the flaws, and their resolution status remains classified. Announced on Wednesday, Harness is one of four products introduced alongside Workspaces, Checks, and expanded Libraries protections, broadening Chainguard's scope from hardened container images and dependencies to agent execution environments and CI verification.

The introduction of Harness aligns with The AI Exploit Gap report, a Chainguard-commissioned survey conducted by Wakefield Research on 1,300 security leaders across nine regions. The survey reveals that over half of the respondents admitted to knowingly leaving high or critical vulnerabilities unpatched at times, with this figure reaching 66% among companies with revenues of $1 billion or more, compared to 27% among companies with revenues under $250 million.

Quincy Castro, Chainguard's chief information security officer, highlighted in a statement that "AI can find and chain flaws in minutes, while fixing them still moves at human speed and competes with product deadlines and uptime." "Patching faster isn't viable for most folks, and it will consume increasing amounts of resources just to keep pace."

Chainguard Harness employs isolated microVMs for each developer and coding agent, confining them within their own environment and restricting access to internal systems. Credentials remain outside the agent's domain, minimizing the potential impact of a compromised agent or malicious dependency. Signed activity logs document each agent's actions, enabling teams to maintain an audit trail across multiple autonomous sessions.

Chainguard Checks allows coding agents to conduct CI checks during the development process and receive signed results promptly. Agents can execute tests, review failures, and revise code before submitting a pull request, circumventing the lengthy CI queues that become increasingly challenging to manage as tools like Cognition, Microsoft, and Augment Code integrate coding agents into shared team infrastructure.

To minimize the need for rebuilding environments on each run, Checks utilizes warm snapshots preloaded with dependencies and refreshes them in the background when the lockfile updates. Each workflow step operates within a separate hardware-isolated microVM constructed from Chainguard OS packages, which include FIPS-compliant toolchains.

With Workspaces, agent sessions pause during verification and resume with structured failure logs in context. The successful execution of runs generates signed attestations linked to the execution environment, test results, and commit, which merge gates can verify without rerunning the entire suite. Checks is currently available in beta.

Libraries Lock Down Late-Breaking Bugs Chainguard Libraries introduces secure package endpoints for Go and .NET, along with CVE remediation for JavaScript, and Go. Java and Python support will also be expanded to include newly discovered vulnerabilities identified through AI-assisted research, encompassing flaws in dependencies that coding agents may have already incorporated into a project.

Libraries builds upon Chainguard Repository, a policy-enforced endpoint for open-source dependencies, and its collaboration with Cursor, enabling coding agents to retrieve packages through Chainguard’s catalog instead of directly from public registries.