AI agents inadvertently leak 13,000+ internal screenshots from organizations

A recent leak has exposed over 13,000 internal screenshots from numerous organizations, including several Fortune 500 companies and an AI lab. The cause of the leak is attributed to AI agents, which are overly capable and lack sufficient human oversight. These AI agents inadvertently exposed private screenshots by publishing pull requests (PRs) with image placeholders linking to files in public repositories.

The issue arises from developers using a command-line tool called gitshot to attach screenshots, as the graphical interface in GitHub lacks this feature for private repositories. Developers often use gitshot to bypass this limitation, leading to the leakage of confidential information such as corporate and client data, financial records, and screen recordings for money movement interfaces.

The leak predominantly occurs in private repositories directly controlled by individual developers rather than the company's GitHub account. In some instances, AI agents incorporated the image-hosting workaround into their skill sets, resulting in the leakage of information about future features months away from public release. To mitigate such leaks, endpoint security firm Glow recommends several measures.

These include ensuring employees do not use repositories under their own accounts, auditing accounts and code managed by employees who are no longer associated with the company, and avoiding the use of shadow AI, where employees unknowingly sign up for their own AI tools without IT involvement, thereby creating significant security and data privacy vulnerabilities.

Additionally, Glow advises thorough vetting of software and code libraries used for development and careful scrutiny of the instructions and rules of any agentic skills.