Tag

#role-confusion

1 story tagged role-confusion.

Dumbbell chart showing attack success rate jumps from baseline to chain-of-thought forgery across three models: gpt-oss-20b from 26% to 80%, o4-mini from 2% to 80%, GPT-5 hardened from 2% to 35%. Chain-of-thought forgery bypasses LLM guardrails at 80% peak ASR.

Chain-of-thought forgery bypasses LLM guardrails at 80%

Chain-of-thought forgery fakes an LLM's reasoning style to bypass guardrails. Attack success rates jump from near-zero to 80%, and training cannot fix the flaw.

Lars Cornelissen · Jul 30, 2026