On 6 October 2026 Anthropic merged Project Glasswing and the Cyber Verification Program into one offer with three access tiers, which includes Claude Mythos 5.1. On 30 September Google had introduced Gemini 4 Argon, for now limited to a group of trusted defenders in the Fairwind Program, while OpenAI reserves the reduced-refusal versions of GPT-6 Astra, released on 3 September, for the Red tier of its Daybreak programme. For your company the choice is between two routes: use the models open to everyone or, if you have a team that defends your own systems, apply for vetted access.
Who can use what
| Model | Who can use it | Conditions |
|---|---|---|
| Claude Fable 5.1 | Generally available | According to Anthropic it finds vulnerabilities; developing exploits is excluded |
| Claude Mythos 5.1 | Vetted organisations | Cyber Verification Program; at launch only US organisations |
| Gemini 4 Argon | A group of trusted defenders | Through Fairwind; Google plans a version without cyber guardrails for them and for its own teams |
| GPT-6 Astra | Generally available, standard safeguards | The reduced-refusal version only in Daybreak Red |
The Cyber Verification Program has three tiers, all including Mythos 5.1, Claude Opus 5.5 and Claude Sonnet 5.5. Defense Access is for security teams that defend systems they own, critical infrastructure operators of any size, small security firms, open-source maintainers and researchers with vulnerabilities already reported. Red Team Access adds authorised penetration testing and is reserved for organisations. Specialized Access is for a few verified organisations that test critical systems, reviewed together with the US government.
Fairwind, launched by Google on 2 September, admits in a first phase governments and national cyber authorities, critical infrastructure operators and core technology platforms; participants reserve access for internal security teams and adopt multi-factor authentication. OpenAI's Daybreak has a Blue tier, for qualified security teams and individuals, and a Red tier, with separate approval and for now only for approved business and enterprise organisations.
Prices per million tokens: Fable 5.1 and Astra $10 for input and $50 for output; Argon $2 and $10 at launch, then $4 and $20.
What the vendors say
- Anthropic writes that Mythos 5.1, with its cyber safeguards off, has the strongest cyber capabilities of any model it has released.
- Google writes that Argon can find, validate and patch critical vulnerabilities on its own, and cites a critical flaw in healthcare software, found in a Wiz initiative, that earlier models had missed; it does not say which software it is or whether the flaw has been fixed.
- OpenAI writes that Astra, with the right tools and access, can find unknown flaws and develop exploits for them across many well-protected systems without a person guiding each step: it is its first model at the Critical level of the Preparedness Framework.
On impact, Anthropic writes that Glasswing partners found at least 129,000 verified vulnerabilities between April and July 2026 and that its own open-source scanning added 5,500 between April and October; more than 33,000 of the total are rated critical or high so far. The data are partial: they come from a subset of partners.
GLM-5.3 is a separate case: an open-weight model from Zhipu AI, released without meaningful safeguards. According to NIST's CAISI, as cited by Anthropic, it is the most cyber-capable open-weight model so far, about four months behind the American frontier. With the smaller GLM-5.3-Flash a researcher built an exploit chain for two already known flaws, including CVE-2026-11645 in Chrome: 20 minutes of human attention, eight hours of model work, $20.40 at Zhipu's API prices.
What your company can do today
Without vetting you can use the models open to everyone for secure code review, threat modelling, patching known issues, finding flaws in your own code and triaging alerts: the list is Anthropic's, which warns that other activities, such as malware analysis, may be interrupted by its classifiers. If you use Google Cloud, Google writes that any customer can use CodeMender with the public models on the Gemini Enterprise Agent Platform.
If you have a team that defends your own systems you can apply for vetted access. Anthropic asks for a description of the work and an attestation of the security controls of the chosen tier, aims to reply within seven business days, requires data retention to spot misuse, and gives Defense Access until 15 December 2026 to adopt phishing-resistant multi-factor authentication. OpenAI reviews requests and states that approval is not automatic: a contract or a purchase alone does not guarantee it.
We recommend treating a model's output as a lead to verify, giving it read-only access to the code with no production credentials, and having a person and the tests review every fix.
Where we start
The news touches two of our services: custom software and management systems and cloud platforms. We deliver code that is tested and documented, and maintenance after delivery is part of how we work. If you ask us for a review, we start from three questions: which components and external services the software uses, who has administrative access and with which permissions, and how long passes between a supplier's fix and its release to production.
Reviewing a fix, even one proposed by a model, means understanding what it makes possible: TrueUp, our GitHub app, computes for each unit of code what it reads, what it writes and which calls it can reach, so whoever reviews a pull request sees which capabilities the change introduces. We cannot guarantee you access to the vendors' programmes, which vet applicants directly, and we do not promise that a model will find every flaw in your software.
If you want a review of this kind on your management system or cloud platform, get in touch.
What we do not know yet
- Google has given a launch price for Argon and no date for opening it to everyone: it writes that this will happen as soon as possible.
- For an Italian company it is unclear whether Mythos 5.1 is available today. On 1 September Anthropic described it as available only to US organisations; the October pages on the programme list no eligible countries and name the organisation's legal and regulatory environment among the criteria.
- According to Anthropic, fewer than half of its partners reported how many flaws were fixed.
- Daybreak's conditions may have changed since 25 September, the date of the archived copy of the OpenAI help page.
Sources
- Anthropic, Expanding the Cyber Verification Program (6 October 2026)
- Anthropic Help Center, Cyber Verification Program (updated October 2026)
- Anthropic, Introducing Claude Fable 5.1 and Claude Mythos 5.1 (September 2026)
- Anthropic, GLM-5.3 and the spread of advanced cyber capabilities (29 September 2026)
- Google, Gemini 4 Argon: our next era of frontier intelligence (30 September 2026)
- Google, Google's Fairwind Program: Cyber defense tools for trusted partners (2 September 2026)
- OpenAI Deployment Safety Hub, GPT-6 Astra System Card (3 September 2026)
- OpenAI Help Center, OpenAI Daybreak - Trusted Access for Cyber Overview
- OpenAI API documentation, GPT-6 Astra Model
