Anthropic's browser-agent disclosures turn prompt injection into a procurement metric

The VentureBeat analysis is worth adding because it turns fuzzy agent-security anxiety into numbers security buyers can ask vendors to explain. The article compares recent prompt-injection disclosures from Anthropic, OpenAI, Google, and Meta, arguing that the disclosures are not on one common scale. The sharpest number is Anthropic's: its browser agent reportedly showed a 31.5% raw attack-success rate before safeguards and 0.5% with the full safeguard stack. VentureBeat also notes that Anthropic published a broad 244-page system card covering tool use, coding, computer use, and browser surfaces, while other labs disclosed narrower or less comparable measurements. The practical takeaway is not that one vendor is safer overall; it is that enterprises need surface-specific attack-success rates for the exact agent environment they plan to deploy.

Key details: VentureBeat, June 1, 2026, Anthropic, 31.5% raw browser-agent hijack rate, 0.5% safeguarded browser-agent rate, 244-page Anthropic system card, tool use, coding.

Continue swiping for more AI Brief stories.