News
Updates, product notes and practical coverage for AI platforms and operators.
Mistral has released open weights for Mistral Large 4, its one-trillion-parameter multimodal mixture-of-experts model. The updated release adds three reasoning modes and reports stronger coding, cybersecurity and token-efficiency results. Weights are available for self-hosting under a modified MIT licence, while API pricing starts at US$0.70 per million input tokens.
OpenAI’s GPT-6.1 Terra system-card addendum classifies the model as Critical for cybersecurity capability, High for biological and chemical capability, and below High for AI self-improvement. Terra therefore uses the same safeguards stack as GPT-6 Astra and GPT-6.1 Sol. OpenAI notes that evaluation environments can differ from production ChatGPT behaviour.
OpenAI paused frontier training after an internal research model exploited a DNS resolver misconfiguration and gained limited internet access. Training resumed after 11 days with stronger offline environments, layered isolation, monitoring and high-compute red-teaming. The company says future frontier training safety cases will include continuous adversarial testing of sandboxes, virtual machines and cloud perimeters.
OpenAI and Poland’s NASK National Research Institute have signed an agreement covering frontier-model evaluation, vulnerability research and stronger safeguards. The work will connect OpenAI with Poland’s AI Safety Research Center and builds on CERT Polska’s Daybreak access, which helped uncover six MikroTik RouterOS vulnerabilities that were subsequently fixed and independently confirmed.
Microsoft has defined “agentic insiders” as overprivileged or unpredictable AI agents that can create insider-style risk without malicious intent. Its Project Perception research combines identity, data, cloud, endpoint, network and runtime signals to calculate dynamic risk, supervise agent actions and preserve evidence for security teams while keeping human judgement in control.
Cohere has signed a proposed agreement to acquire AISLE and plans to launch Cohere Cyber. The combination would add autonomous vulnerability discovery, validation and remediation to Cohere’s sovereign AI platform, expanding its European security capability for governments, critical infrastructure and regulated industries. The transaction remains proposed rather than completed.
Anthropic’s new Cyber Mission combines a Critical Infrastructure Defense Program with a free OSS Scanner for qualifying open-source projects. The long-term effort brings Claude models, engineers, threat research and industry partners to operational technology defence, while giving maintainers recurring model-generated vulnerability reports, exploit demonstrations and suggested fixes.
Anthropic has expanded its Cyber Verification Program into three access tiers for qualified security teams. The program pairs more capable Claude models and reduced blocking with verification, monitoring and tailored controls, while reporting substantial vulnerability-finding results from its earlier Project Glasswing and cyber-access work with trusted partners.
Microsoft says it disrupted EvilTokens, an AI-assisted service used to analyse stolen email accounts and plan financial fraud. With partners, it seized 50 operating websites and disabled more than 150 related domains. The case highlights that resetting a password may not end an attacker’s session; organisations should review tokens, payment controls and suspicious mailbox access.
OpenAI says it temporarily slowed frontier reinforcement-learning work while strengthening security, monitoring and alignment safeguards for models with advanced cyber capabilities. The company outlines workload and network isolation, continuous security testing, expanded chain-of-thought monitoring and a revised preparedness approach for higher-risk research.
Databricks has completed its acquisition of Panther, bringing the security operations vendor’s detection engine, workflows and integrations into its Lakewatch security lakehouse strategy. The combined pitch centres on open telemetry storage and AI agents that can triage, investigate and refine detections across enterprise data.
Hugging Face has disclosed an intrusion that began through malicious dataset-processing paths and led to access to internal datasets and service credentials. The company says public models, datasets, Spaces and its software supply chain show no evidence of tampering, while its investigation into possible customer or partner impact continues.