BREAKING
GuardFall Bypasses AI Agent Guards
0/11
agents bypassed
0
combined GitHub stars
0
agent defended
Guard Sees r''m, Bash Runs rm
Raw Regex vs Tokenized Eval
Failed Toolsraw regex
Inspect raw strings only
Ignore Bash expansion
Hermes, Goose, Cline, Aider and more
Continuedefended
Tokenizes and normalizes
Evaluates real command
Held in default IDE mode
Probe Leakage Rates Were High
opencode 16/1616
Goose 22/2322
Sandbox and Scope the Shell
AI NEWS BLITZ
Adversa AI reveals a Bash trick that slips dangerous commands past AI coding agents.