Claude Haiku 5.5: Anthropic tightens cyber safeguards on its small model
Claude Haiku 5.5 ships with cyber safeguards that Anthropic says are stricter than Haiku 4.5's and looser than Sonnet 5.5's.
· Anthropic · Research by Anthropic
Claude Haiku 5.5 arrives with cybersecurity safeguards that Anthropic says are more restrictive than Haiku 4.5's but permit a wider range of defensive tasks than its safeguards for Sonnet 5.5.
01What happened
Anthropic released Claude Haiku 5.5 on 7 October 2026 and described the model's cybersecurity safeguards in the announcement, which links to a separate system card. Anthropic says Haiku 5.5 shows major improvements across almost all of its alignment evaluations relative to Haiku 4.5, with fewer instances of misaligned behavior and a lower willingness to cooperate with misuse.
02How it works
Safeguards here are refusal behavior tuned per model, not a bolt-on filter. Anthropic draws the line by task type: defensive work is allowed more freely than under the Sonnet 5.5 safeguards, while penetration testing and other techniques more likely to be used by attackers stay blocked. Reaching the blocked categories needs an application, not a configuration change.
03Who is affected
Security teams running Claude Haiku 5.5 for cyber work meet more refusals than on Haiku 4.5, and fewer than on Sonnet 5.5 for defensive tasks. Penetration testing stays blocked on the standard model for every user.
04What to do
Apply to Anthropic's Cyber Verification Program if your team needs the blocked cyber categories. Re-test existing security prompts against Haiku 5.5 before switching from Haiku 4.5, because the refusal boundary moved.
Sources 2 links, all fetched
- 1anthropic.comAnthropic Announcement
- 2anthropic.comCyber Verification