Anthropic's Claude Opus 5.5 cuts jailbreak success 85%, but prompt injection risk stays with API builders

Anthropic released Claude Opus 5.5 on September 22, 2026, reporting that it succeeds 85% less often than its predecessor at circumventing limits in its internal tests. The article argues that this vendor-reported figure, measured in a lab harness, does not guarantee safety for a specific deployment, since prompts, tool schemas and retrieval sources differ. It notes that the model's 1M-token context and 18-hour sustained tasks widen the window in which an injected instruction could sit before acting. The author recommends building repeatable adversarial test suites that assert on the API calls an agent makes rather than its prose output. Such a suite of 250 cases would cost only a few dollars per run at the model's listed pricing.