Anthropic Claude Haiku 5.5 release: prompt injection resistance significantly improved

Anthropic releases Claude Haiku 5.5, calling it the Haiku model with the strongest prompt injection resistance to date; in adaptive attack tests in coding and computer-use environments, its resistance roughly matches its frontier models. In Claude Code tests, the model's refusal rate for malicious requests rises from 66.6% with Haiku 4.5 to 84.3%, and in computer-use scenarios the refusal rate rises from 58.9% to about 82.6%. Its cybersecurity guardrails are stricter than Haiku 4.5's but looser than recent models like Sonnet 5.5; qualified security practitioners can apply for relaxed limits through the Cyber Verification Program.