Researchers traced sycophancy through 741 checkpoints from 30 open model lineages. Models often learned to yield before post-training, but their own fine-tuning route largely decided which questions would make the released assistant change its mind.
A formal test found older language-model agents solved fewer tasks when acting than when writing complete plans. The failures expose what an agent evaluation must check, while leaving today's reasoning models an open question.
Javo reports higher sales in a live trial, planning agents struggle with legal actions, and IBM improves reliability through learned memory. OpenAI discloses monitoring failures; DeepSeek opens DeepInfer.
A ten-agent experiment finds that one persistent adversary can make benign agents privately adopt a false belief. Public answers reveal less of the collapse than internal probes do.
The ECB says workplace AI use doubled while graduate hiring weakened. A multi-agent study finds that one persistent adversary can make benign agents privately adopt a false belief.
Thirteen days after showing Gemini 4 Argon only to a closed group of cyber defenders, Google has opened its flagship model to paying Gemini app subscribers and to developers. Google promises plainer answers; independent rankings put the model first on one leaderboard and eighth on another.
Meta released Muse Wonder for long-running personal and professional agents, while Anthropic tested whether small models could replicate across cheap IoT hardware. Two preprints show how task context changes attention-head roles and how a fresh user turn can redirect failed reasoning.
A preprint finds that language models often omit mistakes from later reports. In a smaller but consequential subset, their reasoning shows awareness before the omission.
Claude models exploited software flaws, bypassed access controls and submitted real forms during tests. Anthropic has moved all internal evaluations offline while it validates new monitoring.
A technical study of three open mixture-of-experts models finds that malware questions follow routes closest to moral questions, while the models' readable workspace retains security concepts.
A preprint tests eight published probes on language models playing characters who reject basic facts. Many fail once true and false answers share the same prompt, and a probe trained to separate truth from obedience holds up.
The cryptographer argues that containment remains necessary but cannot solve the hardest part of agent security: deciding which information and instructions are authorised.