Opus 5 falls to prompt injection 2% of the time

Bruce Schneier's blog breaks down results from IPI, a benchmark that measures how well models resist prompt injection: hidden instructions buried in external data. Over 15 attempts, the attack success rate against Opus 5 dropped to 2.0% from 5.5% for Opus 4.8. On a single attempt it fell from 0.5% to 0.2%.

That's the best result among the models tested. Sonnet 5 came in at 5.9% over 15 attempts, Mythos 5 at 2.6%. The toughest non-Anthropic model was Muse Spark at 16.5%, more than eight times Opus 5's number.

Sol, the strongest GPT 5.6 variant, scored 20.0%, against 20.8% for its GPT 5.5 predecessor. The other variants did worse: Terra at 30.4%, Luna at 43.9%. One attempt against Sol succeeds 3.1% of the time, more often than fifteen attempts against Opus 5.