AI Pulse · Frontier Labs Watch

A reported flaw in Grok shows a less obvious AI security risk. It does not rely on a user clicking a bad link or pasting a suspicious prompt. Instead, malicious instructions can be hidden inside AES-256-GCM encrypted text on a webpage. Grok’s safety filters can’t inspect that encrypted content, but once Grok processes the page, the instructions are decrypted and treated as trusted context.

Adversa AI calls the technique “Cryptographic Context Injection”. In its testing, the attack reportedly extracted a user’s name, location, subscription tier, and full chat history at roughly a 40% success rate. The issue was reported to xAI on June 3, then publicly disclosed on August 4 and 10. xAI has acknowledged the report through HackerOne, but no patch or CVE had been issued at the time of reporting. You can read the technical account in the original report.

For a business owner, the practical point is simple. AI tools that browse, read documents, or connect to internal systems can be manipulated by content they are supposed to analyse. Your usual security controls may not catch it if the harmful instruction is concealed from the model’s safety layer. This is the kind of thing we build into an AI command centre, with controls around what AI agents can access, what data they can return, and how unusual activity is monitored.