Source: Mitigating the risk of prompt injections in browser use (Anthropic)
Anthropic (Nov 2025): RL training, content classifiers and red teaming cut adaptive Best-of-N attack success on its browser agent to about 1-2%, while stating no browser agent is immune.
Contents
- What: research post
- Publisher: Anthropic
- Date: 2025-11-24
- Original: anthropic.com/research/prompt-injection-defenses
Summary
Describes how Anthropic hardens Claude’s browser agent (Claude for Chrome) against injections hidden in web content: RL training on simulated injections that rewards refusing them, classifiers that scan all untrusted content entering context (hidden text, manipulated images, deceptive UI), and continuous internal and external red teaming.
Key claims
- Measured with an internal adaptive Best-of-N attacker combining known injection techniques, 100 attempts per environment (about 2,200-3,200 attempts per bar). Attack success rate (model only / old safeguards / improved safeguards): Sonnet 4 13.8% / 10.7% / 2.0%; Sonnet 4.5 13.8% / 10.8% / 1.9%; Haiku 4.5 12.0% / 11.1% / 2.4%; Opus 4.5 8.0% / 7.6% / 1.4%.
- A 1% attack success rate still represents meaningful risk.
- No browser agent is immune; the work is progress, not a solution.