Anthropic redeployed its strongest public model after a 19-day pause, routing suspicious coding requests to Opus 4.8 instead of simply refusing them.
The new classifier targets a jailbreak first described by Amazon researchers and is said to block more than 99% of those attempts. Anthropic warns the larger safety margin will create more false positives for ordinary developers, making the release a trade-off between smoother coding and stronger protection.
Sources: CocoLoop、Anthropic; MarkTechPost; Infosecurity Magazine