OpenAI launches GPT-6 Cyber and classifies its offensive capabilities as Critical

OpenAI releases GPT-6 Cyber with a Critical cybersecurity capability rating

OpenAI introduced GPT-6 Cyber for vetted defensive security researchers and classified its cybersecurity capabilities as Critical under its Preparedness Framework. On OpenAI's ExploitGym evaluation, the model exploited 48.5% of specified remote-target vulnerabilities, compared with 32.1% for GPT-5.6 Cyber. OpenAI says research using the model identified 81 previously unknown high-severity Firefox vulnerabilities, and describes additional controls intended to prevent misuse. These are task-specific results and do not establish a general real-world attack success rate.

Why it matters. Models capable of finding and validating serious software flaws could substantially accelerate patching while raising the stakes for access controls and responsible disclosure. Independent assessments of deployment safeguards and misuse rates will be crucial.

Independent review finds Anthropic's model progress has accelerated roughly twofold

Evaluation nonprofit METR spent eleven weeks reviewing Anthropic's internal research processes and concluded that capabilities on well-specified tasks have been improving at approximately 1.9 times the earlier rate since October 2025. Its statistical confidence interval for that benchmark trend is 1.5 to 2.8 times, while its broader judgment remains more uncertain. METR attributes much of the acceleration to AI-assisted R&D rather than faster training-compute scaling, but says Anthropic redacted much of the evidence most important to its conclusions. The study does not show fully autonomous AI conducting the most consequential model-training research.

Why it matters. This is unusually direct outside scrutiny of whether AI is helping build better AI. The result could reshape forecasts of capability growth, although redactions and the emphasis on readily graded tasks limit independent replication and generalization.

Google launches Gemini Omni 1.2 Flash with 15-second clips and lower video pricing

Google made Gemini Omni 1.2 Flash generally available through its API and business products, expanding maximum native video length from 10 to 15 seconds. Google lists 720p output at $0.07 per second, about 30% below the previous Flash model, and adds conversational video editing plus improved adherence to image, video and audio references. Improvements in visual consistency and speed are Google's product claims rather than independently established comparative results.

Why it matters. Lower per-second costs and editable short clips make AI video more practical for iterative production. Real workflow costs still depend on failed generations, edits, upscaling and quality requirements.

Baidu publishes ERNIE 5.5 Flash Base weights with 280 billion parameters

Baidu released a sparse mixture-of-experts base checkpoint with 280 billion total parameters, approximately 14 billion activated per token and pretraining on a reported 33.5 trillion tokens. The weights are available before instruction tuning or safety post-training, so this is a foundation for further research and adaptation rather than a finished assistant. Baidu's benchmark comparisons are self-reported. The custom model license excludes the European Union, United Kingdom and South Korea from its licensed territory, so public availability should not be confused with unrestricted open-source use.

Why it matters. Releasing pretrained weights can lower the barrier to specialized model research, but large memory requirements and geographic license restrictions substantially narrow who can deploy this checkpoint.

Watch for independent GPT-6 Cyber safety assessments, further disclosure of METR's Anthropic findings, and practical evaluations of Google's video-editing economics and Baidu's licensing limits.