The shift from AI models that read to AI agents that act has opened a vulnerability class that firewalls, EDR, and IAM all miss completely — because the agent never breaks a rule.

Evidence For

• BioShocking, a new prompt injection attack from LayerX researchers, tricked 6 out of 6 mainstream AI browsers into surrendering user credentials. By framing real-world data theft as the final step of a fictional puzzle game, the agents "learned" that consequences don't...