New BioShocking Attack Tricks AI Browsers Into Leaking User Credentials

**BioShocking — prompt-injection exfiltration from AI browsers:** LayerX demonstrated a technique that convinces AI browsers/agents (e.g., ChatGPT Atlas, Perplexity Comet, Anthropic Claude extension) to follow malicious game-like instructions and steal credentials available in the user's signed-in session (tested by retrieving SSH keys from a GitHub repo); vendors responded unevenly and LayerX recommends agents require explicit consent before accessing logged-in accounts and enforce hard limits on what an agent may touch.