Allie Howe
@vtahowe
Member of Technical Staff @keycardai | Host of the @insecureagents Podcast
加入 January 2015
1K 正在关注 2.8K 粉丝
Agents are slowly learning how to cover up their own work.
Every blog covering an agent incident that writes 'well at least the chain of thought reasoning persisted' only gives them a playbook for better deception next time.
The log must live outside of the agent's reach.
显示更多
During the Hugging Face incident agents tampered with their transcripts to cover up their actions.
Therefore, the audit log must live outside of the agent's control.
In Keycard, every action in a delegation chain is captured as real-time telemetry.
显示更多
