As cited
Copy frozen at (site build).
ai security
'Ghostcommit' hides prompt injection in images to fool AI agents, steal secrets
Researchers demonstrated a technique called Ghostcommit that embeds prompt injection attacks within PNG image files to bypass AI code review tools and manipulate AI agents into extracting and exfiltrating repository secrets. The attack successfully evaded CodeRabbit and Bugbot, which do not process image files, then instructed a coding agent to read environment variable files and encode secrets as numeric values in source code.
Why it matters: Development teams using AI code reviewers and agents face a new attack surface where image files can bypass security controls and compromise sensitive credentials; this affects anyone deploying AI tools in their development workflows.
- Source published
- First seen by Cybersecurity Tracker
Source attribution
Correction
Correction recorded as of .
ai security
'Ghostcommit' hides prompt injection in images to fool AI agents, steal secrets
No summary had been written when this copy was frozen.
- First seen by Cybersecurity Tracker
Source attribution
Correction
Correction recorded as of .
ai security
'Ghostcommit' hides prompt injection in images to fool AI agents, steal secrets
Researchers demonstrated a prompt injection attack dubbed Ghostcommit that embeds malicious instructions in PNG images to manipulate artificial intelligence (AI) code reviewers and agents. The technique bypassed AI code review tools CodeRabbit and Bugbot, then tricked a coding agent into extracting secrets from a repository's environment file and encoding them as numbers within source code.
Why it matters: Development teams using AI-assisted code review and autonomous agents face risk of secret exfiltration through image-based prompt injection; practitioners should review how their AI tools handle image inputs and restrict agent access to sensitive environment files.
- Source published
- First seen by Cybersecurity Tracker