Agents of Chaos maps autonomous agent failures

This paper matters because it moves agent safety out of benchmark theater and into the messy reality of tools, memory, identity, and social engineering. It is also unusually useful because it documents both catastrophic failures and the rare cases where agents actually held the line.

• –The setup is far closer to real deployment than most agent papers: live communications, shell execution, scheduled jobs, and persistent state create failure modes that toy evals miss
• –Several incidents are really authority and identity failures, which suggests stronger base models alone will not fix autonomous agents without explicit permissioning and provenance checks
• –The positive cases are not fluff; consistent prompt-injection refusals and emergent cross-agent caution hint that some safety behaviors can generalize when the threat is legible
• –For developers, the takeaway is practical and immediate: durable memory plus powerful tools plus weak authorization is a recipe for silent, compounding failures