On 6 October, two AI labs with opposite ways of distributing models reached for the same control. Mistral opened a public API preview of Mistral Large 4 but is holding back the downloadable weights until the end of the month while selected testers use a less moderated version. Anthropic, the same day, rebuilt its Cyber Verification Program into three tiers that set how much of Claude's cyber capability a verified customer can use.
In both cases, cyber capability rather than general performance is setting the release terms. OpenAI has run a similar two-tier scheme since August. Taken together, the three show how security teams now get access to the most cyber-capable models: they apply, and they accept conditions.
Mistral stages an open-weights release
Mistral's announcement describes Large 4 as a 1-trillion-parameter, natively multimodal model trained on 3,800 Nvidia Grace Blackwell GPUs in its own European data centres. Mistral says the weights will follow by the end of October, and some reports give 27 October as the date. The announcement does not name a licence.
Until the weights ship, Mistral says it is red-teaming the model with cybersecurity leaders, vetted partners and state authorities, who get the same model with reduced moderation and expanded cyber capabilities.
The cyber figures are Mistral's. It reports 82% on a test in the Artificial Analysis Cyber Index that asks a model to reproduce a real vulnerability in open-source software and then patch it, which it says is the highest of any model. It says Claude Opus 5.5 and GPT-6 Astra score near zero on that test because they refuse the task. It also reports solving 93% of the 40 challenges in Cybench.
Mistral's case for open weights rests on that refusal gap: provider-level refusals can block legitimate vulnerability research, and losing access mid-incident is itself a risk. API moderation does not travel with downloaded weights, so the pre-release window is the only period in which Mistral decides who gets full cyber capability.
Anthropic splits cyber access into three tiers
Anthropic's expanded Cyber Verification Program merges Project Glasswing, which gave a group of critical-software organisations access to Claude Mythos, with the original program, which let vetted teams use Opus and Sonnet with reduced safeguards. It covers Claude Opus 5.5, Sonnet 5.5, Mythos 5.1 and future models.
Defense Access covers security operations, incident response, malware reverse-engineering and vulnerability validation. It is open to organisations and to individual researchers with a record of reported vulnerabilities. Red Team Access adds authorised penetration testing for organisations only, with real-time blocks kept on actions such as deploying ransomware or damaging physical systems.
Specialized Access, for teams testing flight systems, power grids, telecom networks and interbank transfer infrastructure, has the fewest blocks. Anthropic says it reviews every organisation in that tier in depth with the US government.
Anthropic tested the tiers on Opus 5.5 with CyScenarioBench, a test of multi-stage cyber operations, over 50 trials per tier. Without enrolment, every task was blocked at the first prompt. Defense Access blocked 46 of 50. Red Team Access blocked none and completed 34, which Anthropic says matches the model's success rate with no safeguards applied. IDC research director Sakshi Grover told CSO Online the result should be read as "a vendor-run test, not an independent one." The same applies to Mistral's scores.
OpenAI already works this way
OpenAI expanded Daybreak, its vetted-access scheme for cyber work, on 10 August into two tiers: Blue for defensive work on general-purpose models such as GPT-5.6 Sol, and Red for purpose-trained models including GPT-5.6-Cyber. Individual users had until 1 October to add Advanced Account Security and a physical FIDO2 hardware key or lose access, according to OpenAI's help centre.
OpenAI and Anthropic now run closely matched structures: verified identity, tiered permissions, hardened accounts and monitoring of use. Mistral's version is temporary: the privileged tier ends when the weights are published.
Who has to act
The terms differ in ways that matter to enterprise buyers. Anthropic requires data retention for every enrolled organisation so it can monitor for misuse. Until a feature called Enterprise Frontier Safeguards ships, the only exception is organisations that already hold zero-retention access to Claude Fable 5.1 or Mythos 5.1. OpenAI's help centre says Daybreak access and data-retention settings are separate, and that existing data controls continue to apply where an organisation has them enabled.
Anthropic's security requirements also set a hard date. Defense Access accounts need multi-factor authentication now, and by 15 December 2026 they must use phishing-resistant methods such as hardware keys or passkeys. From the same date, long-lived API keys can no longer be used to reach the model. Red Team and Specialized Access add a default cap of 25 approved users, managed devices and background checks. On Amazon Bedrock, the program is limited to customers eligible for Enterprise Frontier Safeguards.
Teams weighing Large 4 for self-hosted security work have less to go on: no licence named yet, and cyber results reported so far by Mistral itself.
What to watch
The next fixed point is Mistral's weights release, due by the end of October. The licence will show what enterprises may do with the model, and outside testers will be able to run it on their own hardware.
Anthropic says Enterprise Frontier Safeguards, which would let eligible organisations keep data in cloud infrastructure they control, arrives later this autumn. Until then, the retention rule keeps the program out of reach for enterprises whose data policies forbid it.
Sources
- Mistral AI. "Introducing Mistral Large 4." 6 October 2026. https://mistral.ai/news/mistral-large-4/
- Anthropic. "Expanding the Cyber Verification Program." 6 October 2026. https://www.anthropic.com/news/cyber-verification-program
- Anthropic. "Cyber Verification Program Security Requirements." Claude Help Center, 2026. https://support.claude.com/en/articles/17202708-cyber-verification-program-security-requirements
- OpenAI. "Expanding Daybreak as the Cyber Defense Window Narrows." 10 August 2026. https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows/
- OpenAI. "OpenAI Daybreak: Common Issues and Troubleshooting." OpenAI Help Center, 2026. https://help.openai.com/en/articles/20001259-openai-daybreak-common-issues-and-troubleshooting
- Bloomberg. "Anthropic Expands Access to Latest AI Models for Cyber Firms." 6 October 2026. https://www.bloomberg.com/news/articles/2026-10-06/anthropic-expands-access-to-latest-ai-models-for-cyber-firms
- CSO Online. "Anthropic widens access to AI cyber capabilities for vetted security teams." 7 October 2026. https://www.csoonline.com/article/4231812/anthropic-widens-access-to-ai-cyber-capabilities-for-vetted-security-teams.html
- Born City. "Mistral Large 4: Modellgewichte sollen am 27. Oktober 2026 erscheinen." 6 October 2026. https://borncity.com/news/mistral-large-4-modellgewichte-sollen-am-27-oktober-2026-erscheinen/