Deep diveAI Cyber Defense
Frontier AI models for cyber defense: the access programs for defenders
Model providers apply cyber safeguards by default, so verified defenders can ask for fewer blocks on legitimate work. This page describes, from the providers' own documents, Anthropic's Cyber Verification Program and OpenAI's Daybreak: their tiers, eligibility, review, data conditions and limits. Terms change often, so check each provider's page before applying.
On this page
- Why providers gate cyber capability at all
- Terms the programs use
- Anthropic's three access tiers side by side
- Anthropic and OpenAI compared on the questions buyers ask
- How an Anthropic application moves from request to use
- Conditions that apply in every tier
- Which route usually fits which kind of defender
- Terms change, and the two providers' pages differ
- Questions and answers
- Sources
Why providers gate cyber capability at all
The same skill that finds a flaw in a library can help someone exploit it. Model providers therefore apply safeguards to cyber requests by default and offer verified defenders a way to ask for fewer blocks on legitimate work. Anthropic states that its generally available models keep conservative cyber safeguards, while its program gives qualifying security professionals access to more capable models with fewer blocks.1
The programs below are described as their providers state them, and the descriptions are not endorsements. OpenAI presents Daybreak as a program for verified defenders.4 Nothing here says that any particular company is enrolled in either program.
Terms the programs use
- Dual-use safeguards
- Classifiers and policies that block or limit requests that could help attackers. Access programs adjust these for verified applicants and leave other safeguards in place.2
- Tier
- A level of access with its own eligibility and controls. Anthropic names three: Defense, Red Team and Specialized.1 OpenAI's Daybreak has Blue and Red tiers.6
- Internal use and external use
- Anthropic treats work on an organization's own code, products and infrastructure as internal use. Work for customers is external use and needs a separate approval.2
- Data retention
- Anthropic requires retention for enrolled users so that it can monitor cyber misuse. Organizations with a zero-data-retention arrangement for certain models can use the program under that arrangement.1
- Trusted Access for Cyber
- The name of OpenAI's verification program, which OpenAI's help center describes as Daybreak. Individual and enterprise applicants follow different routes.6
- Know-your-business verification
- An organization-level identity check. OpenAI's help center says enterprise applicants complete it through a third party before OpenAI decides on eligibility.6
Anthropic's three access tiers side by side
Anthropic's Cyber Verification Program, announced on October 6, 2026, merged its earlier program with Project Glasswing.1 Each tier includes Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1, plus future models.1
| Item | Defense Access | Red Team Access | Specialized Access |
|---|---|---|---|
| Typical work | Security operations, incident response, malware analysis, detection engineering and vulnerability triage and validation | Adds authorized penetration testing, red teaming and exploit development | Authorized red teaming of critical infrastructure and research into the highest-severity techniques |
| Who can apply | Organizations and individuals | Organizations only | Organizations only; running critical infrastructure alone does not qualify |
| Seat limit | None for organizations | Up to 75.2 | Up to 25.2 |
| Review target | A few days | A few weeks, with Defense Access granted meanwhile | Longer, reviewed in depth with the US government |
| Added controls | Multi-factor authentication and frequent key rotation; from December 15, 2026, phishing-resistant authentication and no long-lived API keys.2 | Phishing-resistant authentication, organization-controlled devices, outbound network controls and background checks | All Red Team controls plus single sign-on and managed devices with malware protection |
Seat limits and review targets come from Anthropic's support page, and plan availability for the upper tiers was updated on November 18, 2026.12
Anthropic and OpenAI compared on the questions buyers ask
| Question | Anthropic Cyber Verification Program | OpenAI Daybreak |
|---|---|---|
| Tiers | Defense, Red Team and Specialized.1 | Blue for defensive work and Red for authorized offensive-style testing.6 |
| How eligibility is decided | Applicants are verified and must show the security controls their tier requires. A denied applicant can reapply after seven days.12 | OpenAI reviews each request case by case and says approval is not automatic.6 |
| Individuals | Defense Access is open to individuals who pass an identity check.2 | Applicants must be adults on an eligible paid plan with advanced account security and a hardware security key.6 |
| Organizations | One application per organization. Red Team and Specialized are for organizations only.12 | Organization-level intake with business verification. Red eligibility is not assumed from Blue.6 |
| Security vendors and partners | Partners build the models into products that return defined artifacts. End users do not prompt the model directly.3 | Partners reach models through Daybreak Access, and model access stays with the partner, not the customer.5 |
| Funding for under-resourced defenders | A fund of $35M in Claude credits for organizations helping open-source maintainers.3 | A commitment of $1B in subsidized access, training and technical support for frontline defenders.7 |
How an Anthropic application moves from request to use
- Apply from the right account
Approval does not carry over between organizations, and each organization submits one application.2
- Verification
Anthropic checks identity or business registration and the controls the tier requires.2
- Tier granted
A lower tier can be granted while a higher one is under review.2
- Admin assigns access
Approval does not switch the program on by itself. Administrators assign it to workspaces or roles.2
- Use within scope
Internal use is covered. Work for customers needs a separate approval.2
- Ongoing review
Anthropic may review, narrow or withdraw access at its discretion.2
Conditions that apply in every tier
Anthropic lists work that stays prohibited in every tier: mass destruction of data or devices, denial-of-service attacks, physical-world damage, extortion, election interference and any request with stated malicious intent.2 Its Red Team tier still blocks real-time actions that could cause physical harm or mass disruption, such as deploying ransomware.1
Data conditions matter for regulated organizations. Anthropic requires retention for enrolled users, and organizations that already have zero data retention for certain models can use the program under that arrangement.1 For Amazon Bedrock, Anthropic states that the program is available only to customers eligible for its Enterprise Frontier Safeguards.1
OpenAI's help center says that its Red tier needs separate approval, stronger verification and access controls, and that its newest cyber model is initially limited to approved Red organizations.6 In both programs, approval applies to the named applicant. Staff at another company, a client or a supplier are not covered by it.
Which route usually fits which kind of defender
These are readings of the providers' stated terms, not advice on any one application.
- If
An organization's own staff run security operations and vulnerability work on its own systems.
- If
A consultancy or product company wants to use the models for its customers' security work.
ThenIt needs the provider's separate approval for external use, or it works through a provider-approved partner product.
Anthropic requires separate approval for work done for customers, and states that serving customers without it violates the terms.2
- If
A buyer wants findings from a security vendor that uses these models.
- If
A team does authorized penetration testing or red teaming.
Terms change, and the two providers' pages differ
Questions and answers
Which AI models can verified defenders use for cyber work?
Anthropic states that each Cyber Verification Program tier includes Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1, plus future models.1 OpenAI's help center describes Daybreak Blue as reducing refusals on supported mainline models, with a separate cyber model initially limited to approved Red organizations.6
Can any security team join these programs?
Does joining mean prompts and outputs are retained?
Retention is a condition at Anthropic, which requires it for enrolled users so that it can monitor misuse, with a zero-data-retention route for organizations that already hold that arrangement for certain models.1 Check OpenAI's current terms for its position, because they differ by plan and route.
Does a consultancy need its own approval to use the models for clients?
What does a company get when it buys from a vendor in a partner network?
Sources
- Anthropic: Cyber Verification Program — Anthropic · checked 19 November 2026
- Claude Help Center: Cyber Verification Program — Anthropic · checked 19 November 2026
- Anthropic: Bringing Claude Mythos 5 to more defenders — Anthropic · checked 19 November 2026
- OpenAI: Daybreak — OpenAI · checked 19 November 2026
- OpenAI: Become a Daybreak partner — OpenAI · checked 19 November 2026
- OpenAI Help Center: OpenAI Daybreak, Trusted Access for Cyber overview — OpenAI · checked 19 November 2026
- OpenAI: Daybreak for Frontline Defenders — OpenAI · checked 19 November 2026
More in AI Cyber Defense
Back to AI Cyber DefenseYour starting point
What would you like to move forward?
Start with the outcome that feels closest. We'll help give the next step a useful shape.
Let's find your next move
One conversation. A clearer direction.
Tell us what you want to improve. A few sentences are enough to start.
shayan@coldai.org