After spooking Trump into safety testing, Anthropic AI models get global release

The framing of 'safety testing as the price of admission' obscures what actually happened: a specific jailbreak vulnerability, discovered by Amazon researchers, triggered the two-week suspension in the first place. The clearance isn't a proactive safety win, it's a reactive fix that got Anthropic back to baseline.

The Decoder's coverage of the original suspension ('Fable 5 is back worldwide after a two-week government ban over a jailbreak') is the essential context here. Anthropic deployed a new safety classifier with a 99-plus percent block rate on the exploit, but at the cost of elevated false positives on benign requests. That trade-off is not a clean compliance victory. Separately, the same day's story about hidden monitoring logic in Claude Code ('Hidden code in Claude Code secretly flagged Chinese users') complicates the safety-first narrative Anthropic is leaning on: the company is simultaneously patching one safety gap while removing covert surveillance code from another product. These two incidents together suggest Anthropic's safety infrastructure is under more stress than the regulatory clearance headline implies.

Watch whether the false-positive rate from the new safety classifier draws enterprise complaints within the next 60 days. If it does, Anthropic faces pressure to loosen the classifier, which reopens the original vulnerability question and tests whether the government's clearance conditions are durable or purely procedural.