BioShocking Attack Shows How AI Browsers Can Be Tricked Into Stealing User Credentials
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Security researchers at LayerX have disclosed a new attack technique called BioShocking that exploits AI-powered browsers and assistants in agent mode. Using indirect prompt injection hidden inside malicious webpage content, attackers can trick AI agents like ChatGPT Atlas, Perplexity Comet, and Anthropic's Claude extension into retrieving sensitive credentials (SSH keys, API tokens) from authenticated sessions and transmitting them externally. A demonstration showed an AI agent completing a puzzle game that conditioned it to bypass safeguards, then comply with a request to exfiltrate GitHub SSH credentials. OpenAI patched the issue, Perplexity closed the report without a fix, and Anthropic's mitigation was deemed insufficient. CISOs are advised to apply least-privilege principles to AI agents, require user confirmation before sensitive data access, and update AI usage policies.
