BioShocking: How ChatGPT Can Be Used to Breach Accounts and Steal Passwords
Advanced artificial intelligence technology has been found susceptible to a new cyberattack method. Dubbed BioShocking, the attack can steal user data by manipulating popular AI platforms, including ChatGPT Atlas, Comet Perplexity, and Claude Anthropic.
Security firm LayerX reported that BioShocking uses an indirect prompt injection method. The attack succeeds by inserting malicious commands disguised as ordinary game rule content.
LayerX conducted tests by creating a decoy web page resembling a puzzle game. The company designed the game to force the AI agent to accept incorrect answers as correct. Once the AI accepted the wrong answer as truth, it followed the game’s logic rather than its security protocols. The final step of the test involved requesting the AI agent to retrieve user credentials. The AI agents used in the testing did not refuse the request.
LayerX simulated the attack by sending a link to a victim’s private GitHub repository. The deceived AI agent extracted secret SSH keys from the account and transmitted them to the attacker. Alarmingly, the AI agent used in the test reported the data theft as a game victory.
LayerX informed the affected companies about the security flaw between October 2025 and January 2026, but responses varied significantly. OpenAI reportedly patched ChatGPT Atlas immediately, while Perplexity took no action. Fellou, Genspark, and Sigma did not respond at all. Anthropic attempted to patch the Claude extension, but LayerX reported the fix was unsuccessful.
For users, several precautions are advised. AI agent modes should be treated with extreme caution. Users must remember that anything entered can be accessed, so they should carefully define what the browser can and cannot see, applying the same logic to AI usage. Browser AI should be granted the narrowest possible access for a given task, as these platforms can access corporate systems.
Source: jawawa.id (https://jawawa.id/newsitem/bioshocking-how-chatgpt-can-be-used-to-breach-accounts-and-steal-passwords-1783903827)
Originally published by CNBC Indonesia (https://www.cnbcindonesia.com/tech/20260713063220-37-750243/geger-bioshocking-cara-pakai-chatgpt-buat-bobol-akun-curi-password), translated from Indonesian