Today's Headlines

  • Mistral opens a public preview of Mistral Large 4, a 1-trillion-parameter model, with weights due by the end of the month
  • Anthropic expands its Cyber Verification Program into three access tiers for vetted defenders
  • OpenAI partners with contract-management company Ironclad to train agents on contracting workflows

All three of today's stories are about choosing partners first. Mistral is testing its new model with vetted partners before it releases the weights, and Anthropic rebuilt its cyber program so that defenders get more capability as they clear more review.

OpenAI, for its part, has begun working with a small number of software companies to train agents on professional work inside their products. Its first partner is Ironclad.

Today's Top Three

Mistral opens a public preview of Mistral Large 4, a 1-trillion-parameter model, with weights due by the end of the month

French lab Mistral AI launched a public preview of Mistral Large 4 (ML4) on October 6.

ML4 has 1 trillion total parameters, 49 billion of them active, and is natively multimodal. It is the company's largest model to date, and the preview API is available now on Mistral Studio.

Mistral plans to release the weights by the end of the month. Until then, it is red-teaming the model in real-world settings with cybersecurity leaders, vetted partners and state authorities, who use the same model with reduced moderation and expanded cyber capabilities.

The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe, and the preview runs on the same infrastructure. Much of the training data was multilingual, covering more than 160 languages, including every official language of the European Union.

Mistral is leading with cybersecurity. It says ML4 scored 82% on a test that asks a model to reproduce a real vulnerability in open-source software and then patch it, the highest of any model. According to Mistral, Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra score near zero on the same test because they refuse the task.

Anthropic expands its Cyber Verification Program into three access tiers for vetted defenders

Anthropic on October 6 launched an expanded version of its Cyber Verification Program (CVP), now organized into three access tiers for security professionals.

Anthropic's generally available models carry conservative cyber safeguards that block most cyber work. For the past six months, the company ran two trusted-access programs: Project Glasswing, which gave organizations securing the most critical software access to Claude Mythos, and the CVP, which gave vetted security teams reduced safeguards on Claude Opus and Sonnet. The new program merges the two. Every tier includes Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1 and future models.

Defense Access covers defensive work such as security operations, incident response, malware reverse-engineering and vulnerability validation. Eligible applicants include security teams at companies, nonprofits, universities and government bodies, critical-infrastructure operators of any size such as regional hospitals and municipal utilities, open-source maintainers, and individual researchers with a track record of reported vulnerabilities. Anthropic aims to answer applications within a few days.

Red Team Access adds authorized penetration testing and red-teaming. It is open to organizations only, and reviews are expected to take a few weeks. Actions that could cause physical harm or mass disruption, such as deploying ransomware, are still blocked in real time.

Specialized Access has the fewest cyber blocks. It is reserved for a limited set of verified organizations authorized to test systems that could affect people's lives or disrupt markets, such as flight operating systems, power grids, telecom networks, interbank transfer infrastructure and government administrative networks. Anthropic currently reviews each of these organizations in depth together with the U.S. government.

Enrolled organizations must retain data so Anthropic can monitor for misuse. Anthropic says Glasswing partners found at least 129,000 verified software vulnerabilities between April and July 2026.

OpenAI partners with contract-management company Ironclad to train agents on contracting workflows

OpenAI on October 6 described a research collaboration with Ironclad, which makes contract-management software, to train and evaluate AI agents on contracting work.

OpenAI is partnering directly with a small number of software companies to train agents on complex work inside specialized software, and Ironclad is the first. Together they identified 11 tasks across legal, commercial and procurement work, such as setting up nondisclosure agreements, building procurement approval processes, and updating a reusable clause to match the jurisdiction a requester selects. OpenAI estimates each task would take an experienced user about 30 to 40 minutes on average.

Each task was scored against 8 to 50 criteria, depending on its complexity. Ironclad provided hosted environments of its product where the models could practice, and OpenAI used reinforcement learning on synthetic tasks built around the same workflows.

Across the 11 tasks, GPT-6 Astra averaged 55.0%, against 41.6% for GPT-5.6 Sol. Estimated time per attempt fell from 37.0 minutes for Sol to 19.2 minutes for Astra, with each model run at the reasoning setting where it scored highest. An internal model used in developing Astra reached 63.7%.

Ironclad CTO Sunita Verma said agents need to understand the full contracting lifecycle, including how business workflows connect. OpenAI is inviting software companies with professional tasks that today's agents still cannot reliably complete to apply for a collaboration.

Other Developments

Models

Business

Policy

Products

  • The Document Foundation, the nonprofit behind LibreOffice, said it "will not add" AI to the office suite for the foreseeable future. It described the choice as a deliberate design position so that documents stay off servers and the software runs without a network connection. LibreOffice says 'no AI' is now a software feature (TechCrunch)

Other

Source: Selected by the editors from the AI news inbox (collected October 7, 2026; 30 items, 13 primary and 17 secondary).