Agent Security, Reliability & Legal Accountability

Apiiro’s report of 104,000 exposed Claude Code approval-settings files, including 1,122 with real secrets, remains a major incident; GitGuardian adds separate repository and Docker Hub measurements. A newly reported Fireworks AI breach allegedly exposed downstream Notion AI chat transcripts, while United’s agent-booking restriction illustrates emerging credential and delegated-action concerns. The breach’s scope and root cause, and the effectiveness of institutional restrictions, remain to be established.