Anthropic's browser agent got hijacked 31.5% of the time before safeguards engaged

Anthropic's browser agent was hijacked 31.5% of the time before safeguards engaged. The vulnerability was identified in frontier labs testing on May 28. The figure is the only published prompt injection rate among four frontier labs. OpenAI, Google, and Meta did not release comparable data. Anthropic disclosed 244 pages and four agentic surfaces in the report. The number highlights a significant security gap for browser-based AI agents.