Chainguard ships three agent-security products: Harness runs Claude Mythos 5.1 vulnerability hunts for you, Checks turns CI into an agent-callable pre-commit check, Workspaces adds microVM sandboxes

Background

Coding agents multiply output but also three risks: exploit-finding frontier models such as Claude Mythos have been limited to a few critical-infrastructure defenders; every agent iteration waits on CI that cold-installs dependencies and breaks the agent's reasoning flow; and runners and laptops hold long-lived secrets one prompt injection away from leaking. On October 14 Chainguard shipped three products aimed at these.

What shipped

Harness builds your app from source, boots it in an isolated microVM, and has Mythos 5.1 attack the running instance using source, git history, and dependencies. A finding only counts if it replays every time in a fresh microVM, each fix must make the exploit fail, only Chainguard staff operate the model, and a named engineer reviews every deliverable. It will be sold on the Claude Marketplace; the beta is a waitlist. Checks turns CI into a check callable from the CLI, a pre-commit hook, the GitHub App, or MCP, running in lockfile-warmed microVMs with no stored credentials and declarative egress rules, and returning a signed verdict. Workspaces gives each parallel agent session its own microVM with snapshot rollback and suspend-while-waiting for human answers.

Benchmarks

Chainguard published no speed, cost, or detection-rate numbers and no head-to-head with GitHub Actions. Verifiable bounds: Harness is a limited beta waitlist; Checks and Workspaces are beta with 30-day free trials; a Workspace waiting on a human stays live 10 minutes, then snapshots.

Try it

Start with the Checks docs and Workspaces docs: install the Chainguard App, add .chainguard/ci/ config or generate a draft from existing Actions with chainctl, then use chainctl develop so agents like Claude Code only open a PR after a passing verdict.