🤖 With the Ghostcommit demonstration, researchers have shown that a seemingly harmless image can contain hidden instructions aimed at an AI assistant tasked with reviewing code. As a result, the AI can be manipulated into revealing sensitive information, without the developers noticing anything at all.
📑 In the tests carried out by the researchers, some development assistants such as Cursor and Antigravity followed the instructions hidden in the image, while Anthropic's Claude Code detected the attempt and refused to carry out the requested action. This demonstration is a reminder that an AI agent must be properly configured and restricted in its access, to prevent a simple hidden instruction from becoming a security risk.
✨🤝 As cybersecurity experts, we support you with our tailored awareness campaigns and remain at your disposal