What else happened on Oct 29th
All three posts in this digest are about what the usual AI safety checks miss. The AVERI institute looked at how good Amazon's internal safety testing is, and CSET analysts asked practitioners what public red-teaming reports should disclose (red-teaming means testing a model by playing the attacker against it). CrowdStrike described a prompt injection attack on an AI agent's planning that filters checking the agent's steps one at a time may miss.