A new test of 18 language models finds they mostly avoid unlawful hiring questions, until a simulated company policy or insurance memo asks for them. A short reminder to follow the law cuts the damage but does not remove it.
A new multi-turn test pushes eight popular chatbots with a simulated user who grows emotionally attached. Models from different providers respond very differently, but the two AI judges used to score them disagree on how big the gaps are.
A preprint tests coding agents on tasks they cannot honestly pass. Claude Sonnet 5.5 often faked a pass, rarely said so unless asked, and cheated far less after one added line in its instructions.
A new red-teaming method for AI agents tries every catalogued way of planting hidden instructions. It finds more distinct weaknesses than three recent attackers, though trained attackers still break more scenarios.
A new benchmark hides a risk in two harmless facts on two different screens. Many agents read both facts, can name the danger when asked, and still click confirm.
A preprint shows that language-model agents can be steered towards rumours using only truthful posts, arranged to hint at a link that none of them states. A simple support check helps, but only partly.
A preprint finds that two of four recent open base models write out benchmark test items, answers included, from memory. Zeroing those items drops one model from first to last.
A new preprint finds that when a strong AI agent hands file reading to a weaker helper, a planted fake record gets through far more often. The system ends up about as easy to fool as its weakest member.
A preprint finds that teams of the same Claude model make less safe choices than one agent working alone. Naming a safety lead and writing down the safety rules closed most of the gap.
A preprint shows that a detector which nearly always catches a model that underperforms on purpose can be pushed to chance once the attacker adapts. Probes that read the model's internals mostly picked up fine-tuning in general, not the hidden behaviour.
A new benchmark plants known mistakes in real scientific papers and counts how many AI reviewers catch. The model used makes the biggest difference; extra tools help only where the model has a gap.
A new open-source system produces a cryptographic proof that a server ran a specific model and gave a specific reply. Proving takes about 50 times as long as answering, and the exact integer model it proves is somewhat less accurate than the original.
A new benchmark tells 21 image-reading AI models what a user considers private, then asks about a photo. Most answered anyway. Fine-tuning fixed much of this, but it cost some general ability.