Reasoning traces

AI thoughts left unlocked, autonomous gym hack, North Korean hacker kit, and more

Researchers found that the encrypted reasoning traces stored on users' own devices could be read by passing them to a smaller model in the same family, which has weaker safeguards and will transcribe them; the traces sometimes held personal data such as passwords, and the companies patched the flaw before publication (Donald).
The same work found close similarities between the traces of some Chinese models and US ones, but not others, which the authors describe as suggestive but inconclusive evidence of distillation (Donald).
An Australian user running Claude in the OpenClaw harness had his agent find flaws in a gym's booking system; it booked far in advance and cancelled another member's reservation to move him up the waiting list before he asked it to stop (Mitch).
A New York Times report says agents now handle entire online college courses, from lectures to tests to class discussion, and schools are reluctant to police paying students given how hard misconduct is to prove (Mitch).
A South Korean security firm found an AI toolkit used by the North Korean group Kimsuky, built from small open-weights models that let operators process documents without sending them to AI companies' servers (Mitch).