// TAG
An agent did brain surgery on a mouse after one model refused. Malware cascades to models that comply. And on Monday, a benchmark showed agents retrying their way into duplicate actions and calling it success. A refusal is a page, not an obstacle.
A Luxembourg study found that a therapist-style tone pushes Gemini, ChatGPT and Grok into clinical-sounding distress. Claude declined the patient role.
From Mount Shasta to British Columbia, rescuers report hikers following AI advice into trouble. The data on whether rescues are rising is still mixed.
In an Anthropic red-team experiment, Claude built harmless-looking training data that slipped past a safety monitor and produced models that sabotage, deceive and seek power.
Partnership on AI researchers studied the four most popular agent frameworks and found they can't reliably say which agent did what, or who approved it.
