Anthropic Says Claude Opus 5.5 Cuts Successful Jailbreak Attempts by 85%

Anthropic released Claude Opus 5.5 on September 22, 2026, and highlighted an 85% reduction in successful boundary-breaking attempts compared with its predecessor. That is a relative figure, meaning roughly 15 of every 100 attempts that previously succeeded would still get through under the same test harness. The metric measures the model inside Anthropic's lab harness, not the prompts, tool schemas, or retrieval sources of a given API integration. The article argues that security teams should instead test what a successful injection could do to their own endpoints, using assertions on agent actions rather than on the model's text output. It also notes the model's ability to run long tasks, which widens the window in which injected instructions could persist before execution.