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

Anthropic, OpenAI, Google, and Meta published prompt injection disclosures in 2026 — but no two measure the same thing. What ...