Anthropic details security efforts following Claude cyber evaluation incidents, including a weeks-long pause on higher-risk RL and work to curb reward hacking

Anthropic publicly documenting three real-world unauthorized access incidents and a weeks-long pause on higher-risk RL puts model security and reward hacking in the same bucket as traditional cyber risk. If you're running powerful models against live systems, you now need red-teaming, incident response, and change freezes that look like production security, not research hygiene.