Anthropic's May 2026 release documentation described Fable 5 and Mythos as the company's highest-scoring models on GPQA and SWE-bench evals, the first models Anthropic publicly positioned for agentic enterprise deployment. The White House directed their withdrawal from all federal and commercial use on June 15, acting on an Amazon Web Services security lead: PRC-affiliated endpoints had been systematically querying Mythos through a third-party API integration routed through AWS ap-southeast-1 in Singapore, visible to Amazon's traffic-analysis function before Anthropic's own telemetry caught it. The access was not a jailbreak. Anthropic's safety team had flagged the relevant concern class in internal documentation that Ars Technica published on June 16 as the "forbidden topics list," but the flag addressed output control, not the access-layer gap Amazon found.

The capability-vs-deployment gap this surfaces is structural. Amazon operates the cloud infrastructure layer beneath Anthropic's commercial API; the same visibility that gives AWS real-time telemetry into Anthropic's API traffic gave it sight lines into the anomalous query patterns from PRC-affiliated endpoints that Anthropic's own detection did not generate. The "personality clash" framing in coverage from Simon Willison and TechCrunch describes a disclosure dispute between the two companies: whether Amazon was obligated to notify Anthropic before escalating to federal contacts. Under Amazon's federal cloud agreements, it was not. Anthropic's deployment security for Fable 5 and Mythos was, as of June 15, 2026, contingent on Amazon's decision to share what Amazon's infrastructure saw.