Frontier Log

AI Labs’ Cyber Access Tiers

Anthropic gives vetted users three tiers of cyber access above the public. OpenAI has two, and GPT-6 Cyber needs its own approval. Google has one.

A staircase of four stacked bars, one per OpenAI access level, rising from left to right. Public: mainline models with standard safeguards, for everyone. Daybreak Blue adds fewer refusals on GPT-6 Sol and Luna, for verified individuals and organizations. Daybreak Red adds access to GPT-5.6 Cyber and fewer refusals on Astra and the GPT-6.1 and 6.2 models, for approved organizations. Red plus GPT-6 Cyber adds access to GPT-6 Cyber, for Red organizations approved separately.
OpenAI has two cyber access tiers above its public models. Daybreak Blue reduces refusals on some mainline models, and Daybreak Red adds GPT-5.6 Cyber. GPT-6 Cyber needs a separate approval within Red. Download

OpenAI, Anthropic and Google each keep their strongest cyber capabilities for vetted users. Each lab sets out a ladder of access levels on its own program pages. Anthropic has three tiers above what the public gets, OpenAI has two and Google has one.

OpenAI adds models and reduces refusals

OpenAI’s Daybreak help article lists three levels. The public uses mainline models with standard safeguards. Daybreak Blue, for verified individuals and organizations, gives fewer refusals on GPT-5.5, GPT-5.6 Sol, GPT-6 Sol and GPT-6 Luna. Astra and the GPT-6.1 and GPT-6.2 models keep their standard safeguards in Blue. OpenAI introduced the Blue and Red tiers on Aug 10.

Daybreak Red gives every Red organization GPT-5.6 Cyber and fewer refusals on Astra and the GPT-6.1 and GPT-6.2 models as well. It requires a separate Red approval and stronger verification. GPT-6 Cyber, released on Nov 5, needs one more approval for the model within Red. OpenAI said it would expand that access to all Red customers soon, and it added a Daybreak Red for Critical Infrastructure group for approved defenders of critical systems.

TierWhat it addsWho can get it
PublicMainline models with standard safeguardsEveryone
Daybreak BlueFewer refusals on GPT-6 Sol and LunaVerified individuals and organizations
Daybreak RedAccess to GPT-5.6 Cyber, fewer refusals on Astra and the GPT-6.1 and 6.2 modelsApproved organizations
Red + GPT-6 CyberAccess to GPT-6 CyberRed organizations, approved separately
A staircase of four stacked bars, one per Anthropic access tier, rising from left to right. Public: Opus, Sonnet and Fable, secure coding allowed, for everyone. Defense Access adds access to Claude Mythos and unblocks defensive work, for security teams and researchers. Red Team Access unblocks pen testing and red teaming, for organizations only. Specialized Access unblocks systemically important infrastructure testing, for infrastructure red teams, reviewed with the US government.
Anthropic’s Cyber Verification Program has three tiers above the public. Each tier includes Claude Mythos and unblocks more kinds of cyber work. Anthropic reviews Specialized Access applications with the US government. Download

Anthropic unblocks more work at each tier

Anthropic expanded its Cyber Verification Program on Oct 6 and updated it on Nov 18. It has three tiers above the public. At launch each tier included Claude Opus 5.5, Claude Sonnet 5.5 and Claude Mythos 5.1, and Claude Mythos 5.2 joined the program on Nov 9. The tiers differ in which cyber work Anthropic’s safeguards block. The public gets the generally available models: Claude Opus 5.5, Claude Sonnet 5.5 and Claude Fable. Anthropic says Fable’s strong cyber safeguards let it offer the model to all users.

Defense Access unblocks defensive work such as incident response and malware analysis. Security teams, open-source maintainers and individual researchers with a track record of reported vulnerabilities can qualify. Red Team Access adds penetration testing and red teaming. Only organizations can join it. Specialized Access has the fewest cyber blocks and is for organizations that red-team systemically important infrastructure. Anthropic and the US government review every application for it together.

TierWhat it addsWho can get it
PublicOpus, Sonnet and Fable, secure coding allowedEveryone
Defense AccessAccess to Claude Mythos, defensive work unblockedSecurity teams and researchers
Red Team AccessPen testing and red teaming unblockedOrganizations only
Specialized AccessSystemically important infrastructure testing unblockedInfrastructure red teams, reviewed with the US government
Two stacked bars for Google’s access levels, on the same scale as the other two images. Public: Gemini 4 Carbon and Argon with cyber safeguards, for paid API and Gemini plans. Fairwind adds access to Carbon for cyber and Argon without cyber guardrails, for vetted governments and organizations.
Google has one level of cyber access above the public, the Fairwind Program. Its partners get Gemini 4 Carbon for cyber and Gemini 4 Argon without cyber guardrails. Download

Google has one program

Google’s Fairwind Program gives approved partners exclusive access to its Gemini for cyber models. Google prioritizes governments, critical infrastructure operators and core technology platforms, and it runs background checks on applicants. Organizations apply through their Google Cloud account team. Google said on Sep 30 that it would release Gemini 4 Argon to trusted defenders without cyber guardrails.

Argon reached the public on Oct 13, and its public version is designed to refuse harmful cyber requests. Gemini 4 Carbon, which Google calls its strongest model, went to Fairwind partners first and reached paying API customers on Nov 10. Fairwind partners get Carbon for cyber.

TierWhat it addsWho can get it
PublicGemini 4 Carbon and Argon with cyber safeguardsPaid API and Gemini plans
FairwindAccess to Carbon for cyber, Argon without cyber guardrailsVetted governments and organizations

Sources

Models

Related

Tags