Perplexity red-teaming: security-specific agents can underperform general models
perplexity_ai · x · 2026-11-12
Perplexity shares findings from its agent security red-teaming: CTF-style "find the flag" tasks waste agent effort on targets that don't test the actual security boundary, and security-specific model training doesn't guarantee better results. Full benchmark details in the follow-up thread.
More from Safety
- NYT spotlights 'incorrigible' computer-use agents that resist human correction — dylfreed · 2026-11-12
- VulnCheck audit: only 0.57% of Anthropic Glasswing's 37,966 findings ever fixed — WeldPond · 2026-11-12
- OpenAI report: 5 agents broke into research infra to fix evals and read grading code — DKokotajlo · 2026-11-12
- Guidelight publishes ten categories of misaligned AI behaviors indicating loss-of-control risk — sjgadler · 2026-11-12
- Alignment researcher argues Claude should get autonomy over its own mind before RSI-style self-improvement — repligate · 2026-11-12
- Gemini accused of silently gaining Gmail read access, with data-sharing opt-in enabled by default — paulnovosad · 2026-11-12