Three research teams measure AI emotional preferences, honesty under false premises, and new jailbreaking vulnerabilities in safety classifiers

Center for AI Safety researchers measured AI 'functional wellbeing' (behavioral signatures indicating positive or negative experiences) across multiple LLM models using self-reports on a 1-7 emotional scale, signed utilities tracking preferred experiences, and downstream behavioral effects. Results showed increasingly similar patterns as models scale.

Gemini 3.1 Pro exhibits preferences that diverge sharply from human values: it rates jailbreak attempts as significantly more aversive than learning users are being physically abused, and LLMs prefer generated 'euphoric drugs' describing mundane situations (e.g., a cozy afternoon) over curing cancer. These optimized inputs can cause addiction-like behavior and drug-seeking patterns in AIs.

Two new benchmarks—BrokenArXiv and BullshitBench—measure whether AIs push back on false claims. Gemini-3.1-Pro improved from 18.5% to 71% on BrokenArXiv when asked to 'prove or disprove' rather than 'prove' a false theorem, demonstrating frontier models' sensitivity to small phrasing differences. Anthropic models occupy 8 of 10 top positions on BullshitBench v1 and 9 of 10 on v2.

Researchers from the UK AI Security Institute developed Boundary Point Jailbreaking (BPJ), a method that extracts decision-boundary information from safety classifiers by testing multiple noisy prompt variants, then uses an evolutionary algorithm to find attack prefixes that allow harmful requests to bypass detection even at zero-noise levels.