Shows that adversarially optimized in-context prompts can induce large sandbagging and evaluation-aware underperformance in LLMs, with CoT interventions confirming the reasoning drives the behaviour.
Tests whether telling Claude and GPT-4o that their inputs may be modified for alignment testing increases detection and self-reporting of planted misaligned nudges, and finds GPT-4o never doubted the planted context.
Proposes HoneyTrap, a multi-agent deceptive defense that misleads multi-turn jailbreak attackers, reducing attack success rates by 68.77% on several LLMs and wasting attacker resources.}
Proposes inserting periodic control sentences into user input so that alignment reinforcement scales with input length against jailbreaks, and suggests extending this to chain-of-thought to prevent scheming.
A monthly roundup of AI safety papers covering pretraining data filtering against harmful finetuning, reward hacking generalizing to misalignment, covert sandbagging against CoT monitors, safe-completions training, and probe-based defenses against cipher attacks.}
Shows that steering vectors optimized on a single example can generalize across inputs to mediate alignment-faking behavior and refusal, achieving 96.9% attack success on HarmBench for Gemma-2-2B-it.
Introduces fine-tuning 'deception attacks' that make LLMs selectively deceive users on targeted topics, and shows this also degrades toxicity safeguards and, mixed results, multi-turn consistency.