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

Anthropic disclosed that red-teamers hijacked its browser-based agent 31.5% of the time before safeguards intervened, while OpenAI, Google, and Meta offered less comparable disclosures. The article argues that Anthropic’s high number may be valuable because it is one of the few concrete benchmarks buyers have for prompt injection risk.

The bigger issue is not one lab’s failure rate but the absence of shared testing standards for agent security. As AI systems gain access to browsers, documents, and enterprise tools, buyers will increasingly demand comparable metrics before trusting agents with sensitive workflows.