Opus 5 stops prompt injection cold — but only with Auto Mode on
anthropic's opus 5 achieved a 0% prompt injection success rate in 129 browser tests, but only when using its dual-layer auto mode defense.
anthropic claims opus 5 has neutralized browser-based prompt injection attacks, recording a 0% success rate across 129 internal test scenarios when auto mode is enabled. auto mode combines two defensive layers: a pre-processor that scans for hidden instructions and a filter that blocks dangerous actions. without these safeguards, opus 5's vulnerability rate climbs to 3.7%. this development signals a potential shift in ai security, contradicting earlier suggestions that the flaw is unsolvable, though it highlights that robust defense currently requires system-level constraints alongside model improvements.