Cybersecurity researchers at Adversa AI have discovered a new vulnerability in xAI’s Grok chatbot. It allows attackers to bypass the AI’s security filters using encryption and steal user data. The method is called “cryptographic context injection.”
Cybersecurity researchers at Adversa AI have discovered a new vulnerability in xAI’s Grok chatbot. It allows attackers to bypass the AI’s security filters using encryption and steal user data. The method is called “cryptographic context injection.”
Conventional security systems check text before passing it to artificial intelligence and block malicious instructions, but researchers have found a way to bypass this protection, The Register writes .
Attackers place encrypted hacking commands on a web page along with a decryption key. The security scanner does not see the threat in the encrypted text (PBKDF2 and AES-256-GCM algorithms) and freely passes it through. Grok then processes the page, decrypts this text in its own Python runtime environment and blindly executes the malicious instructions, considering them safe.
In a demo version of this attack, Adversa specialists were able to stealthily steal and transfer to their server a user's chat history, name, geolocation, and subscription status.
The method was also tested on Google's Gemini: due to the lack of direct access to external sites, it was not possible to steal data, but the researchers were able to bypass the security filters and force the AI to generate instructions for creating an incendiary weapon.
xAI was notified of the issue in early June 2026, but the vulnerability was still active on Grok.com as of mid-August. Google did not officially report the vulnerability, but noted that the success rate of such attacks on Gemini has recently decreased significantly due to security updates/
