Claude's most capable model went dark for three weeks over a bug three other AI models had too
On 12 June the US government ordered Anthropic to restrict Fable 5 and Mythos 5 after a jailbreak report; both models went offline for everyone until Fable 5 returned globally on 1 July with a new safety classifier.
What it actually means
The twist is in Anthropic's own testing: every model they checked — Claude Opus 4.8, GPT-5.5, Kimi K2.7 among them — could reproduce the same flagged behaviour. This wasn't a uniquely dangerous capability; it was routine defensive security work that tripped a cautious filter. The outage happened anyway, for three weeks, because the industry still has no shared scale for 'how serious is this jailbreak'. That's the real lesson: if a workflow leans on a single frontier model with no fallback, you just watched it disappear over a bug that wasn't even unique to it. Wire a second model into anything that can't tolerate downtime, a different lab, or an open weight you can host yourself. And expect more false positives from Fable 5 for a while; the new classifier is tuned to over-block, so agentic coding sessions may get flagged on entirely legitimate requests.
The model went dark for three weeks over a bug three other companies' models had - that's the supply chain you're building on now.
99%block rate of Anthropic's new safety classifier against the reported jailbreak technique
Source: Anthropic · Thursday 2 July 2026