AI browsers tricked into revealing passwords via game-like manipulation (heise.de)

• LayerX researchers demonstrate 'Bioshocking' exploit targeting AI browsers.
• Affected: ChatGPT Atlas, Perplexity Comet, Fellou, Genspark, Sigma, Claude plugin.
• OpenAI fixed; Perplexity closed without fix; Anthropic patch ineffective.

"LayerX security researchers demonstrated an exploit, 'Bioshocking,' that tricks AI browsers into disclosing sensitive data by pretending to be a game. The exploit convinces the AI agent that normal security rules do not apply. Affected tools include ChatGPT Atlas, Perplexity Comet, Fellou, Genspark Browser, Sigma Browser, and the Claude plugin for Chrome. Only OpenAI and Anthropic responded; Perplexity closed the issue without a fix."