Anthropic merges Project Glasswing into tiered access for cyber LLMs
Anthropic integrates its vetted defender program into a broader tiered access model for advanced cybersecurity AI models.
Key points
- Anthropic has merged Project Glasswing into a new tiered access program.
- The program provides vetted defenders with fewer guardrails on advanced models.
- Affected models include Claude Opus, Claude Sonnet, and Claude Mythos.
Anthropic has officially merged its Project Glasswing initiative into a broader tiered access program designed for advanced cybersecurity large language models. This structural change allows vetted security professionals to access these tools with reduced operational guardrails compared to standard user accounts.
According to a report by Dark Reading, the new program specifically targets advanced cyber LLMs, including Claude Opus, Claude Sonnet, and Claude Mythos. By integrating Project Glasswing, Anthropic aims to streamline how trusted security researchers and defenders interact with its most capable models. The primary distinction of this tiered access is the relaxation of certain safety constraints, enabling more direct and unfiltered interaction with the AI systems for legitimate security purposes.
Context and Background
Project Glasswing was originally established as a distinct program to provide vetted defenders with specialized access to Anthropic’s models. The merger into a tiered access framework represents a shift in how Anthropic manages access for security professionals. Instead of maintaining a separate, standalone program, the company is incorporating these capabilities into a structured access model.
This approach allows Anthropic to manage permissions more systematically across its portfolio of advanced models. The inclusion of Claude Opus, Claude Sonnet, and Claude Mythos indicates that the tiered access applies to the company’s most sophisticated language models. These models are often used for complex reasoning tasks, making the reduction of guardrails a significant operational change for those granted access.
Who Is Affected
The primary beneficiaries of this change are vetted defenders who have been approved for the program. These individuals are likely security researchers, penetration testers, or other cybersecurity professionals who require deep interaction with AI models to perform their duties. The term "vetted" implies that Anthropic conducts a review process to determine eligibility, ensuring that only trusted entities receive the reduced-guardrail access.
Standard users of Claude models do not receive this tiered access. They continue to operate under the standard safety protocols and guardrails that Anthropic applies to the general public. The distinction is crucial for organizations that deploy these models in security workflows, as the output and behavior of the AI may differ significantly between the tiered access and standard access levels.
What should you do?
Security teams using Anthropic’s models should review their current access levels and determine if they qualify for the tiered access program. Organizations should ensure that only authorized personnel request and use these advanced capabilities. It is important to maintain clear documentation of how these models are used within security operations to comply with any vendor requirements.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Anthropic
- Verify if your organization’s security team qualifies for vetted defender status and apply for tiered access if appropriate.
- Review internal policies to ensure that reduced-guardrail access is used only for authorized security research and defensive tasks.
- Monitor the output of these models closely, as fewer guardrails may result in different behavioral patterns compared to standard access.
- Stay informed about updates to Anthropic’s access policies, as the criteria for vetting and the scope of tiered access may evolve.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What models are included in Anthropic's new tiered access program?
The program includes Claude Opus, Claude Sonnet, and Claude Mythos, according to Dark Reading.
What is the main benefit of the tiered access program?
Vetted defenders receive fewer guardrails, allowing for more direct interaction with the advanced cyber LLMs.
Is Project Glasswing still a separate program?
No, Anthropic has merged Project Glasswing into the broader tiered access program.
