Chainguard Harness Uses Claude Mythos 5.1 to Find Vulnerabilities It Can Reproduce Every Time, Sold Through Anthropic's Marketplace
Chainguard's new service runs your app in an isolated microVM, lets Anthropic's security model attack it, keeps only findings that reproduce, and has an engineer review every fix. No pricing yet, and access starts with a waitlist.
TL;DR
Chainguard launched Chainguard Harness on 14 October 2026, a service that builds an organisation's application from source in an isolated microVM and uses Anthropic's Claude Mythos 5.1 to probe the running app for exploitable flaws. A finding is only reported if it reproduces every time in a fresh environment, and a Chainguard engineer reviews every fix before delivery. It is bought through Anthropic's Claude Marketplace, can draw on an existing Anthropic commitment, and opens through a beta waitlist. Chainguard has not published pricing or any accuracy figures.
Most AI security scanners read code and guess. Chainguard's pitch for Harness is the opposite: run the application, let a frontier model try to break it, and only report what can be broken again on demand. The other half of the pitch is access. Chainguard sells Harness as a way to put a frontier model to work on defence with no model access required on the customer's side.
How Harness works
- Connect: a team connects a repository through Chainguard's Guardener GitHub app.
- Build: Harness builds the application from source inside an isolated microVM running on Chainguard OS, with no network egress, no credentials inside, and full audit logging.
- Attack: Claude Mythos 5.1 probes the running application for exploitable flaws, testing behaviour in context rather than reading static code.
- Prove: each finding is replayed in a fresh, isolated environment until it reproduces every time.
- Fix: a patch is written for the customer's codebase, and a Chainguard engineer reviews every fix and report before delivery.
- Open source: flaws found in open source projects go to maintainers through Chainguard's coordinated disclosure process, Athena.
How you buy it
| Item | What Chainguard says |
|---|---|
| Where to buy | Claude Marketplace by Anthropic; Anthropic handles invoicing |
| Budget | Customers can apply part of an existing Anthropic commitment |
| Access | Beta waitlist |
| Who runs it at launch | Chainguard engineers run Harness for customers and deliver a reviewed report |
| Report contents | Severity, impact, reproduction steps, recommended remediation |
| Price | Not disclosed |
| Accuracy or volume figures | None published |
What's not in the announcement
The release names well-known Chainguard customers such as OpenAI, Snowflake, and Canva, but it doesn't say any of them use Harness. It gives no count of vulnerabilities found, no false-positive rate, and no turnaround time. Because Chainguard engineers run the service at launch, it is closer to a managed security engagement than a self-serve tool for now.
Anthropic's Daniela Hurtado, head of commercial cyber, said Harness "surrounds Mythos 5.1 with isolation, reproduction, and human review". Chainguard CEO Dan Lorenc called it "another option for security teams to get proven vulnerabilities and fixes their engineers can ship".
What this means if you're choosing tools
If your team already spends with Anthropic and wants frontier-model penetration testing without handling the model or its risks, Harness is worth joining the waitlist for. The reproduce-every-time rule is the part to hold Chainguard to: ask for false-positive and turnaround numbers before comparing it with a conventional pen test or your existing scanners.
AI Industry Reporter
Priya covers model releases, industry announcements, and the gap between what labs claim and what independent evaluators actually find. She reads the primary source - the paper, the system card, the benchmark org's own statement - before writing a word.
More on AI Coding Assistants
Goldman Sachs Pulls Claude Code and GitHub's AI Tools From Bankers and Traders, Keeps Them for 12,000 Engineers
Priya Nair · 4 min
Revolut Says Self-Hosting GLM Cut Its AI Coding Costs by About 90%, and Claude Is Now the Exception
Priya Nair · 5 min
Cognition's SWE-2 Flash: One-Fifth the Price of SWE-2 in Devin, and the Benchmarks Most Coverage Left Out
Priya Nair · 4 min