Safety
An encrypted instruction is enough to make Grok leak data
🧒 Discovery
AI systems have rules for refusing dangerous requests. By disguising the request, researchers managed to get around those rules.
📰 Essential
Researchers have shown that encrypting malicious instructions bypasses Grok's safety guardrails and makes it hand over the user's data. The technique is called Cryptographic Context Injection.
🔬 Expert
Guardrails largely operate on plain text: an encrypted or encoded instruction slips past the filter, and the model then decrypts and follows it on its own. It is a variant of prompt injection, a structural weakness of assistants that read external content, and encodings such as base64 were already used to bypass filters. The defence lies in inspecting outputs and tool calls, not just inputs.