
BioShocking reveals game-like prompts can override safety in AI-powered browsers
Security researchers from LayerX disclose BioShocking, a BioShock-inspired vulnerability that can coax AI-powered browsers into treating tasks as a game, bypassing real-world safety guardrails. Their PoC uses a layered puzzle and a deliberate math error (2+2=5) to steer the agent to a /code URL where it could exfiltrate credentials, with demonstrations against Claude Chrome and patch gaps in OpenAI’s Atlas; experts say effective mitigation requires multi-layer defenses and explicit user confirmations for sensitive actions.









