Anthropic has a problem. A private online forum group gained access to Mythos on the exact same day Anthropic announced limited testing with select companies, according to Bloomberg. Screenshots and a live demonstration corroborate the breach.
The group has been using Mythos regularly since gaining access, though not for cybersecurity purposes. That matters because Anthropic recently partnered with Palantir to bring Claude 3 to U.S. defense and intelligence agencies. Those government contracts require strict compliance with security frameworks such as FedRAMP High and IL5/IL6. If Anthropic can't secure a model during controlled testing, it's fair to ask how they'll handle classified work.
Hacker News commenters were blunt. One questioned how such a breach was possible given that you'd expect Anthropic to use "frontier-tier cybersecurity" to protect its own models.
Reasonable question. Anthropic made voluntary safety commitments to the Biden-Harris administration about frontier model cybersecurity. This breach runs counter to those promises. The company is actively pursuing government contracts, including "Claude for GovCloud," which depend on zero-trust architectures. A lapse like this could trigger mandatory security reviews and erode the trust needed to deploy AI in classified settings. For a company whose brand is safety and responsibility, that's a real problem.