BREAKING
GuardFall Bypasses AI Agent Shell Guards
0
agents tested
0
vulnerable
0
K
combined GitHub stars
Why Text Guards Fail
1
Input r''m looks safe
↓
2
Guard passes the text
↓
3
Bash quote removal yields rm
↓
4
Destructive command runs
Continue Held the Line
10 Other Agents
vulnerable
●
Regex on raw strings
●
Leaky tokenized guards
●
No guard or disabled sandbox
Continue
0 of 21 bypassed
●
Tokenize-and-canonicalize
●
Normalizes commands first
●
Blocked most cases structurally
Guards Can Create False Security
Scope the Shell, Drop Auto-Approval
AI NEWS BLITZ
Security firm Adversa AI has disclosed a technique that defeats shell guards in AI coding agents.