Chainguard unveiled an AI-powered security harness on Wednesday that deploys frontier language models, including Anthropic's restricted Claude Mythos 5.1, to identify and remediate exploitable vulnerabilities in application code that conventional scanning tools fail to catch. The tool allows development teams to leverage Mythos capabilities without direct model access, though Chainguard has not specified which models perform vulnerability detection for customers outside Anthropic's Project Glasswing program.
The company reports that Harness has uncovered serious vulnerabilities in codebases that had already undergone rigorous security assessments, including systems deployed in critical infrastructure environments. However, specifics regarding the affected projects, the nature of the discovered flaws, and their current remediation status remain undisclosed.
The Harness announcement is part of a broader four-product rollout that also includes Workspaces, Checks, and expanded Libraries protections. These offerings extend Chainguard's security footprint beyond hardened container images and dependency management into coding agent execution environments and CI pipeline verification.
Mythos access and marketplace availability
Anthropic has limited Claude Mythos 5.1 access to participants in Project Glasswing, creating uncertainty about which models will power Harness for teams outside that program. The harness will also be offered through Anthropic's Claude Marketplace, enabling organizations to apply existing Anthropic spending commitments rather than allocate separate budget.
Enterprise patching crisis revealed
Chainguard commissioned Wakefield Research to survey 1,300 security leaders across nine markets, producing findings published as The AI Exploit Gap. The research shows that more than half of respondents acknowledge their organizations knowingly leave high or critical vulnerabilities unpatched on occasion. At companies generating $1 billion or more in annual revenue, this figure climbs to 66%, compared with 27% at organizations with revenue below $250 million.
AI can find and chain flaws in minutes, while fixing them still moves at human speed and competes with product deadlines and uptime
Quincy Castro, Chainguard's chief information security officer
Castro elaborated that accelerating patch deployment is not practical for most organizations and will increasingly consume resources simply to maintain current security posture.
Isolated execution for coding agents
Each developer and coding agent operates within a hardware-isolated microVM with network policies that constrain access to internal systems. Credentials remain external to the agent environment, reducing the attack surface if an agent becomes compromised or a malicious dependency is introduced. Signed activity logs track agent actions across multiple autonomous sessions, providing audit trail capabilities.
CI verification within agent workflows
Chainguard Checks enables coding agents to execute CI checks during the development phase and receive signed results within seconds. This capability allows agents to run tests, analyze failures, and iterate on code before submitting pull requests, bypassing the CI queue bottlenecks that emerge as vendors including Cognition, Microsoft, and Augment Code deploy coding agents as shared team infrastructure running concurrent sessions.
To avoid environment reconstruction on each execution, Checks leverages warm snapshots with pre-installed dependencies and refreshes them in the background when the lockfile is updated. Individual workflow steps execute in separate hardware-isolated microVMs built from Chainguard OS packages, incorporating FIPS-compliant toolchains. When Workspaces is enabled, agent sessions pause during verification and resume with structured failure logs available in context.
Runners do not maintain long-lived credentials. Instead, a host-side proxy injects ephemeral credentials into outbound requests, while declarative egress policies govern network access and can terminate a microVM if it attempts a disallowed connection. Successful runs generate signed attestations linked to the execution environment, test results, and commit, allowing merge gates to verify outcomes without re-executing the test suite. Checks is currently available in beta.
Dependency protection and CVE remediation
Chainguard Libraries is introducing secure package endpoints for Go and .NET, along with CVE remediation for JavaScript and Go. Support for Java and Python will expand to address newly disclosed vulnerabilities identified through AI-assisted research, including flaws in dependencies that coding agents may have already incorporated into projects.
These enhancements build on Chainguard Repository, a policy-enforced endpoint for open source dependencies, and a partnership with Cursor that enables coding agents to resolve packages through Chainguard's catalog rather than pulling directly from public registries.