Chainguard announced the launch of Harness, an AI-powered security platform designed to discover and remediate exploitable vulnerabilities in application code, including flaws that traditional security scanning tools may not detect. The platform uses advanced models, including Anthropic’s Claude Mythos 5.1, while details about the models available to customers outside Project Glasswing have not been disclosed.
Chainguard said Harness discovered serious vulnerabilities in repositories that had already undergone extensive security scans, including software used in critical infrastructure. However, the company did not disclose the names of the affected projects, the nature of the flaws, or whether they had been remediated, so the scope or generalizability of the findings cannot be assessed based on the published information.
A Launch Beyond Container Scanning
Harness is the centerpiece of a launch package comprising four products and updates, expanding Chainguard’s scope from secured container images and dependencies to coding-agent execution environments and verification within continuous integration pipelines. The package includes Workspaces, Checks, and expansions to Libraries protection.
Customers can use Harness without direct access to Claude Mythos 5.1. The product will also be available through the Claude Marketplace, allowing teams to use their existing spending commitments with Anthropic rather than creating a separate budget. Chainguard did not specify which models will handle vulnerability discovery for customers not participating in Project Glasswing.
Isolating Coding Agents from Internal Systems
Each developer or coding agent receives a hardware-isolated microVM, with network policies limiting access to internal systems. Credentials remain outside the agent environment, while signed logs record executed activities to provide an audit trail across multiple autonomous sessions.
The environment uses declared egress policies that can terminate a microVM when an unauthorized connection is attempted. Execution environments also do not retain long-term credentials; instead, an agent running on the host injects short-lived credentials into outgoing requests.
Bringing CI Tests into the Agent Workflow
Chainguard Checks allows coding agents to run continuous integration checks during development, then read failure states and modify code before opening a pull request. The service uses prebuilt snapshots containing dependencies to avoid rebuilding the environment for every run, updating them in the background when the lockfile changes.
Each part of the workflow runs inside a separate microVM built from Chainguard OS packages, with FIPS-compliant toolchains. Successful runs produce signed attestations linked to the execution environment, test results, and software commit, allowing merge gates to verify them without rerunning the entire suite. Checks is currently available in beta.
Why Does This Matter?
Chainguard highlights a practical gap between the speed at which AI discovers flaws and the slowness of fixing them. According to a survey commissioned by the company and conducted by Wakefield Research among 1,300 security leaders in nine markets, more than half of organizations said they sometimes leave high or critical vulnerabilities unpatched. The rate rises to 66% among companies with more than $1 billion in revenue, compared with 27% among companies with less than $250 million in revenue.
Chainguard’s explanation is that AI can discover and link flaws within minutes, while remediation remains constrained by product schedules and service continuity. In practice, Harness does not by itself provide an automatic solution to the remediation problem; the source does not explain the success rate of fixes, identify the vulnerabilities discovered, or specify the time required to remediate them. The product’s value therefore remains tied to teams’ ability to review the results and integrate changes safely.
In parallel, Chainguard Libraries is adding secure access points for Go and .NET packages and expanding CVE handling to JavaScript and Go, with plans to expand support for Java and Python. These changes aim to address new flaws, including vulnerabilities in dependencies that coding agents may have already added to projects, relying on Chainguard Repository and its partnership with Cursor.