Anthropic’s Claude Haiku 5.5 Improves Prompt-Injection Resistance and Cyber Safeguards
Summary
Anthropic says Claude Haiku 5.5 is more resistant to malicious prompts and prompt injection than its predecessor, with stronger cyber safeguards. Its offensive-test results remained below Sonnet 5.5 and Opus 5.5.
Key points
- With safeguards disabled, Haiku 5.5 achieved arbitrary code execution in 4 of 410 tests involving known Chrome V8 flaws.
- A pre-release version completed 3.3% of multi-stage cyber-operation challenges, versus 46.1% for Sonnet 5.5 and 67.6% for Opus 5.5.
- In Claude Code tests, it refused 84.3% of malicious requests, up from 66.6% for Haiku 4.5; computer-use refusals rose to 82.6% from 58.9%.
- Anthropic reports improved prompt-injection resistance, though Haiku 5.5 remained less resistant than its larger models on a separate benchmark, particularly in graphical computer use.
- The evaluations excluded some production protections. Anthropic noted remaining weaknesses in sensitive conversations and advised API developers to add safeguards.
- Anthropic says the model’s safeguards block penetration testing and other attacker-prone techniques, while allowing a wider range of defensive tasks; eligible professionals can apply for reduced restrictions.
Article Details
- Event Type
- Anthropic released Claude Haiku 5.5, a new AI model, and reported cybersecurity, safety, and prompt-injection evaluation results.
- Impact
- The model performed better than Haiku 4.5 on reported exploit, code-security, harmful-request refusal, and prompt-injection tests, but remained behind Sonnet 5.5 and Opus 5.5 in some offensive-capability and resistance evaluations. Anthropic reported lower pricing and availability across its platform and major cloud platforms.
People
Vendors
Amazon Web ServicesThe model is available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure.AnthropicAnthropic’s Claude Haiku 5.5 delivers faster performance and stronger cybersecurity capabilities than Haiku 4.5, while still remaining below Sonnet 5.5 and Opus 5.5 in offensive skill.Google CloudThe model is available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure.Microsoft AzureThe model is available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure.
Products
AI TeammatesWe ran it through our eval suite for AI Teammates, our AI agent product, covering use cases like triaging bugs, setting up projects, and searching large portfolios to surface high-risk or overdue work.Claude CodeHaiku 5.5 refused more malicious requests in Claude Code and computer-use tests.Claude Haiku 5.5Anthropic’s Claude Haiku 5.5 delivers faster performance and stronger cybersecurity capabilities than Haiku 4.5, while still remaining below Sonnet 5.5 and Opus 5.5 in offensive skill.Claude PlatformDevelopers can access it on the Claude Platform using the model identifier claude-haiku-5-5.claude.aiWith a near-final version of Claude.ai’s production system prompt applied, Haiku 5.5 achieved a 99.71% harmless response rate on harmful requests.Haiku 4.5Anthropic’s Claude Haiku 5.5 delivers faster performance and stronger cybersecurity capabilities than Haiku 4.5, while still remaining below Sonnet 5.5 and Opus 5.5 in offensive skill.Opus 5.5Anthropic’s Claude Haiku 5.5 delivers faster performance and stronger cybersecurity capabilities than Haiku 4.5, while still remaining below Sonnet 5.5 and Opus 5.5 in offensive skill.Sonnet 5.5Anthropic’s Claude Haiku 5.5 delivers faster performance and stronger cybersecurity capabilities than Haiku 4.5, while still remaining below Sonnet 5.5 and Opus 5.5 in offensive skill.