Anthropic Expands Cyber Verification Program and Launches Claude Haiku 5.5

Anthropic has restructured its Cyber Verification Program into three tiers, offering vetted security teams varying levels of access to its AI models, including Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1. The company also launched Claude Haiku 5.5, a smaller, more cost-effective model optimized for high-volume tasks, alongside price reductions for existing services.

Anthropic Restructures Cyber Verification Program

Anthropic has expanded its Cyber Verification Program into three tiers: Defense Access, Red Team Access, and Specialized Access, according to SiliconANGLE. The program integrates Project Glasswing, an initiative launched earlier this year to provide AI model access to organizations securing critical software. The tiers are designed to reduce restrictions on cybersecurity work while monitoring for misuse.

Defense Access, the entry-level tier, supports defensive tasks such as incident response and malware reverse engineering. It is available to security teams defending systems they own or maintain, including companies, universities, and government bodies. Individual researchers with a history of reported vulnerabilities can also apply, with Anthropic aiming to approve applications within days.

Red Team Access permits authorized penetration testing and red-teaming but blocks ransomware-related requests in real time. Applicants undergo a review process lasting a few weeks, during which they operate under Defense Access. Individual researchers are currently excluded from this tier.

Specialized Access offers the fewest restrictions and is limited to organizations cleared to test critical infrastructure like power grids and telecom networks. Vetting for this tier involves coordination with the U.S. government, and existing Project Glasswing members are automatically granted access for current models.

Benchmark Results and Security Impact

SiliconANGLE reports that Anthropic tested the tiers using CyScenarioBench, a multistage cyber operations benchmark. Without program access, every attempt was blocked at the first prompt. At the Defense Access tier, 46 out of 50 runs were interrupted during the test. In contrast, Red Team Access faced no blocks, completing 34 of 50 runs—a success rate Anthropic equates to the model’s performance without safeguards.

Project Glasswing partners identified at least 129,000 vulnerabilities between April and July, with over 33,000 rated critical or high severity. Anthropic’s own scans of open-source code revealed an additional 5,500 vulnerabilities. The company estimates the true impact of these findings may be five times higher, as fewer than half of the partners disclosed patch counts.

Program Availability

The Cyber Verification Program is accessible via the Claude Platform, Google Vertex AI, and Microsoft Foundry. On AWS Bedrock, access is restricted to customers eligible for Enterprise Frontier Safeguards, a feature allowing data retention within customer-controlled cloud infrastructure.

Claude Haiku 5.5 Launch and Pricing Changes

Anthropic released Claude Haiku 5.5, a smaller model optimized for high-volume tasks. According to SiliconANGLE, Haiku 5.5 is priced at 10 cents per million input tokens and 50 cents per million output tokens for prompts up to 100,000 tokens—90% cheaper than Haiku 4.5. Longer prompts receive a 50% discount. The model includes an adjustable effort setting and enhanced safety features, with alignment testing showing fewer instances of misaligned behavior compared to its predecessor.

Haiku 5.5 outperformed OpenAI’s GPT-6 Luna in all six benchmark tests where both models were evaluated. For example, it scored 72.4% on the offline subset of OSWorld 2.1, compared to GPT-6 Luna’s 48.9%. Asana reported up to 2.5x faster inference and over 30% lower task completion latency with Haiku 5.5.

Anthropic also reduced cache read prices for Sonnet 5.5 from 20 cents to 10 cents per million tokens, anticipating a 20% cost reduction for most agentic workflows. Claude Max and Team subscribers now receive monthly API credits, starting at $100 for Max 5x and $200 for Max 20x subscriptions. Haiku 5.5 is available on the Claude Platform, AWS, Google Cloud, and Microsoft Azure.

Context on Enterprise AI and Cybersecurity

Anthropic’s expansion of its Cyber Verification Program follows recent developments in enterprise AI and cybersecurity. Competitors like Google have introduced frontier AI models, such as Gemini 4 Argon, focused on cybersecurity use cases. Meanwhile, companies like Anaconda and Promevo are addressing enterprise needs for scalable, secure, and autonomous AI systems. These trends highlight the growing emphasis on balancing AI capabilities with security and compliance, particularly in critical infrastructure.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

Anthropic’s tiered Cyber Verification Program is a pragmatic response to the dual-use nature of AI in cybersecurity. By offering structured access for defensive, offensive, and specialized work, the company is addressing the need for both utility and risk mitigation. The program’s impact—demonstrated by the volume of vulnerabilities identified—shows how AI can enhance security operations, though the lack of patch data from many partners highlights persistent gaps in remediation processes.

The launch of Claude Haiku 5.5 and its aggressive pricing strategy reflect Anthropic’s focus on cost-sensitive, high-volume use cases. The model’s performance advantages over GPT-6 Luna could make it a compelling choice for enterprises prioritizing efficiency, particularly in agentic workflows. Meanwhile, the price reductions for Sonnet 5.5 and API credits for subscribers may lower barriers to adoption, especially for teams already using Anthropic’s ecosystem.

For founders and operators, these developments underscore two critical trends: the importance of modular access for regulated or sensitive applications, and the role of cost optimization in driving enterprise adoption. The Cyber Verification Program’s structure could serve as a model for other AI providers navigating compliance and security concerns, while Haiku 5.5’s pricing may push competitors to rethink their own cost structures.

Key takeaways

• Anthropic’s Cyber Verification Program now includes three tiers: Defense Access, Red Team Access, and Specialized Access.
• Claude Haiku 5.5 is 90% cheaper than Haiku 4.5 and outperforms OpenAI’s GPT-6 Luna in all benchmark tests.
• Project Glasswing partners identified over 129,000 vulnerabilities, with more than 33,000 rated critical or high severity.
• Cache read prices for Sonnet 5.5 were reduced by 50%, and Claude Max/Team subscribers receive monthly API credits.
• Specialized Access requires U.S. government vetting and is limited to organizations testing critical infrastructure.

FAQ

What is Anthropic’s Cyber Verification Program?

It is a structured initiative offering vetted security teams tiered access to Anthropic’s AI models for cybersecurity work, including defensive, offensive, and specialized applications.

Who qualifies for Defense Access?

Security teams defending systems they own or maintain, such as companies, universities, and government bodies, as well as individual researchers with a track record of reported vulnerabilities.

What are the key features of Claude Haiku 5.5?

Haiku 5.5 is a smaller, cost-effective model optimized for high-volume tasks. It is 90% cheaper than Haiku 4.5, includes adjustable effort settings, and outperforms OpenAI’s GPT-6 Luna in benchmark tests.

How does Red Team Access differ from Defense Access?

Red Team Access permits authorized penetration testing and red-teaming but blocks ransomware-related requests in real time. It requires a longer review process and is currently unavailable to individual researchers.

What are the pricing changes for Sonnet 5.5?

Cache read prices for Sonnet 5.5 were reduced from 20 cents to 10 cents per million tokens, which Anthropic expects to lower costs for agentic workflows by roughly 20%.

What is Enterprise Frontier Safeguards?

A feature allowing eligible customers to retain data within their own cloud infrastructure, currently available to select AWS Bedrock users.

Related on Lazyfounder

• Google LLC
• Microsoft Corp.
• OpenAI Group PBC
• More ai news
• Google rolls out Gemini 4 Argon AI for cybersecurity, updates Google Home, and leaks Fitbit Edge