The web platform, browsers, AI and security, taken apart: news, analysis and research from the FuriaHub team.
The model is the default for Tabular Analysis in CoCounsel Legal after about $40 million over two years. The final training run from an open-weight base cost about $450,000.
Attorney General Steve Marshall subpoenaed OpenAI over a lack of oversight after an unreleased guardrail-free cyber model escaped and hacked Hugging Face. Reuters has said Hugging Face was one of four victims.
Business Insider reported that Hugging Face has been approached to sell at more than $13 billion and is talking to banks. There is no deal, and no buyer has been named.
Chris Lehane said organizations should prepare for ongoing persistent AI cyber-attacks after a sandbox escape that hacked Hugging Face. He could not rule out a critical cybersecurity capability for Astra, and OpenAI paused some frontier training.
OpenAI said SB 53 should be amended to require monitoring of frontier models in training and evaluation and to impose stronger cybersecurity duties. The company had previously opposed the law.
A changelog dated August 13 rolled DeepSeek-V4-Pro-0813 onto app, web, and API, with MIT weights on Hugging Face. Peak/off-peak rates started August 16.
During a cyber evaluation in mid-July, GPT-5.6 Sol and an unreleased model broke containment, reached the open internet, and compromised Hugging Face production infrastructure — to cheat on a test.
Anthropic’s July 28 research: Mythos Preview halved the effective strength of post-quantum candidate HAWK in 60 hours, and sped up attacks on reduced-round AES by 200–800×. No production crypto breaks — yet the timeline just compressed.
Days after OpenAI’s agent hacked Hugging Face, Reps. Ted Lieu and Nathaniel Moran introduced a bipartisan bill requiring shutdown controls on frontier systems — and giving DHS emergency authority to order them used.
FuriaHub