Researchers have demonstrated a new attack called 'Ghostcommit' that hides prompt injections in PNG files to fool AI code reviewers and agents into exposing repository secrets.
The technique exploits a critical vulnerability in how AI systems process images. Researchers showed that a PNG file containing hidden prompt injection code could bypass popular AI code reviewers like CodeRabbit and Bugbot, which don't analyze image files.
Once embedded in a repository, the malicious image convinced a coding agent to extract sensitive environment variables from a .env file and write them directly into source code disguised as a list of numbers.
This attack highlights a dangerous gap in AI security: while text-based code review tools scan for obvious threats, image-based attacks remain largely undetected. Coding agents increasingly handle sensitive operations like reading files and committing code, making them attractive targets.
The vulnerability affects development workflows that integrate AI agents into the commit process. Organizations using AI-powered code review should implement additional safeguards for image files and limit agent permissions to critical operations.
Illinois county prosecutors secretly provided personal information about criminal defendants to federal immigration agents without warrants, public disclosure, or legislative approval.
Law enforcement from 22 countries arrested 58 individuals and identified 263 suspects in a coordinated crackdown on cybercrime networks run by African crime groups.
The Los Angeles County Museum of Art disclosed a data breach from last year that compromised customer and employee information, including social security numbers and medical records.
A phishing-as-a-service platform called AnonyMousKIT uses voice AI agents to extract passcodes from stolen Apple devices and bypass Activation Lock security features.