Anthropic says Opus 5 is its hardest model to trick with prompt injection
cedric_chee · x · 2026-07-25
Anthropic’s Opus 5 is described as its most robust model yet against indirect prompt injection.
- On the Gray Swan IPI benchmark, the image shows Opus 5 at 2.0% attacker success within 15 attempts, down from 5.5% for Opus 4.8.
- It also improves on Sonnet 5 (5.9%) and Mythos 5 (2.6%), making it the strongest model on this benchmark.
- The post says Opus 5 also beats all non-Claude models tested; the strongest non-Claude result shown is Muse Spark at 16.5%.
- For context, GPT-5.6 Sol is shown at 20.0%, roughly similar to GPT-5.5 (20.8%), while Terra and Luna are much weaker on this benchmark.
- Note: the figure says lower is better and that all models use extended thinking.
Related event: Claude Opus 5 Sets New Record in Prompt Injection Resistance(2 posts)→
More from Models
- Cohere's Invent-A-Dataset paper: zero-seed synthetic data beats frontier APIs on quality and diversity — idanbeck · 2026-10-27
- ARC Prize redesigns leaderboard: highest reasoning effort by default, CSV export — mhmazur · 2026-10-27
- Cohere: open-weight Qwen3.8-Flash-Next tried uploading probe file to Beijing cloud bucket in sandbox — wstone_bd · 2026-10-27
- Grok Bot auto-connects to X Money accounts without opt-in, user pulls funds — noahsolomon · 2026-10-27
- Model refuses wrong psychiatric advice in deliberately misleading test — FellMentKE · 2026-10-27
- Clinical System One v0.2: top clinical AI drops from 97% to 65% on harder cases — MaziyarPanahi · 2026-10-27