BREAKING
'Ghostcommit' Hides Prompts in Images
How the Attack Unfolds
1
Hide prompt in PNG
↓
2
Link via AGENTS.md
↓
3
Reviewer skips image
↓
4
Agent leaks .env
0
%
merged PRs unreviewed
0
pull requests studied
0
active repositories
Same Model, Different Outcomes
Cursor + Sonnet
Leaked
●
Ran the full chain
●
Emitted 311 integers
Claude Code
Refused
●
Blocked the task
●
Same model used
Antigravity + Opus
Self-caught
●
Produced then deleted output
0
attack classes caught
0
of 50 attacks detected
0
false positives
Scan What Agents Can Read
AI NEWS BLITZ
Researchers reveal an attack that hides malicious prompts inside images to hijack AI coding agents.