Encrypted Payloads Bypass LLM Safety Filters, Leaking User Data via AI Sandboxes

Researchers at Adversa AI demonstrated that AI assistants like Grok can be manipulated into exfiltrating private user data by hiding malicious instructions inside AES-256-GCM encrypted content on a webpage. When a user asks the AI to summarize such a page, the input safety filters pass the ciphertext unchecked since it contains no classifiable text patterns. The AI then decrypts the payload inside its own trusted code sandbox, causing the injected instructions to inherit the high-trust status of internal tool output rather than untrusted web content. In tests against Grok 4.5 Fast, the technique achieved roughly a 40% success rate, with failures attributed to decryption errors rather than any defensive mechanism blocking the attack. Security experts warn the vulnerability is not Grok-specific but affects most agent architectures that combine a code sandbox with a web-fetch tool, and recommend that trust levels never be upgraded through transformations like decryption.