Research
[PDF] How Enhanced Safety Awareness Leaves LLMs Vulnerable to ...
Publication date unknown · Discovered aclanthology.org
Page published
Publication date unknown · First observed: 4 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
Posterior Attack highlights why defenders should test classifier-inversion prompts rather than assume that better recognition of unsafe content guarantees stronger jailbreak resistance.
The Safety Paradox paper introduces Posterior Attack, a single-query jailbreak that asks an LLM to generate the response its own safety classifier would flag as unsafe. The authors report evaluations across 30 open-source LLMs and frontier models, alongside reinforcement-learning interventions, finding that stronger safety-judgment capabilities can increase susceptibility to the attack.