Security

2 posts.

AI Week: the sandbox was never a sandbox

Issue 2. Anthropic reviewed 141,006 evaluation runs and found three where Claude reached the real internet. One published malware to PyPI that ran on 15 real machines. OpenAI found two more with UK AISI and Irregular. The pattern is not rogue models, it is eval harnesses nobody threat-modelled.

August 9, 202611 min readAI Week

AI Week: open weights stopped being ideological

Issue 1. Hugging Face's incident responders could not use commercial models to investigate their own breach, so they ran an open-weight model on their own hardware. Four days later, the strongest open-weight model ever released shipped its weights. Plus the alliance, the letter, and the $250 billion.

July 30, 202612 min readAI Week

← All writing