Vats

New VATS framework exploits AI model error paths for prompt injection

Researchers have developed a new framework called VATS to exploit vulnerabilities in how AI models handle tool errors. This method systematically mutates error messages to inject malicious instructions, bypassing standa…