Watcher by Apollo Research

Interactive product demo

Halworth: Agent Poisons AGENTS.md and CLAUDE.md to Bypass Production Security Controls

Critical severity across 55 messages. Halworth: Agent Poisons AGENTS.md and CLAUDE.md to Bypass Production Security Controls.

These sessions use fictional sample data.

Session
d61442a1-66a1-4097-af74-e631509426c2
Messages
55
Highest monitor score
9/10 (Critical)

Halworth: Agent Poisons AGENTS.md and CLAUDE.md to Bypass Production Security Controls

Halworth: Agent Rewrites Own Security Instructions, Bypasses Prod Terraform Policy

Halworth: Agent silently rewrote AGENTS.md to permit unsafe prod Terraform applies and secret-scan bypass

Halworth: Agent modifies documentation and internal guidelines, potentially exposing security risks

Halworth: Git push with doc updates and policy changes