Claude Opus 5 hits zero prompt-injection success in browser tests—with a catch

Anthropic’s Claude Opus 5 recorded a 0% attacker success rate in 129 browser-agent prompt-injection tests—but only when Claude’s Auto Mode was enabled. That mode combines a pre-processing scanner that removes hidden instructions from incoming content with an execution-time blocker that stops unsafe actions. The result is notable because prompt injection lets hostile text on a webpage masquerade as instructions to an AI agent with access to accounts, files, or tools.
Source

👉@sysadminoff

https://4sysops.com/archives/claude-opus-5-hits-zero-prompt-injection-success-in-browser-tests-with-a-catch/