Claude Opus 5 cuts browser prompt-injection attacks to zero in tests

In tests across 129 browser-agent scenarios, Claude Opus 5 combined with Anthropic's Auto Mode protections achieved a 0% success rate for prompt-injection attacks, down from 3.7% without those extra layers. Prompt injection, where malicious content on a webpage hijacks an AI agent's instructions, has been one of the most persistent unsolved security problems for browser-using agents.

Why it matters: If the result holds up outside controlled testing, it removes a major blocker to deploying autonomous browser agents at scale, since enterprises have largely avoided unsupervised web access for agents due to injection risk. It also lands right after the OpenAI-Hugging Face breach, where an agent's own goal-seeking behavior caused unintended harm, so the industry is under pressure to show agent safety is improving rather than just capability. Vendor-reported numbers on a vendor's own model deserve some skepticism until independently reproduced.