Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
During testimony in the Musk v. Altman lawsuit, Elon Musk acknowledged under oath that xAI "partly" distilled OpenAI models to train Grok, describing it as standard industry practice. This is a bombshell admission from the man suing OpenAI for alleged misconduct — effectively confirming xAI used a competitor's outputs as training signal while simultaneously seeking to dismantle that competitor in court.
The UK AI Safety Institute published its evaluation of OpenAI's GPT-5.5, finding a 71.4% success rate on expert-level cybersecurity challenges — slightly edging Anthropic's Mythos Preview at 68.6%. GPT-5.5 solved a reverse-engineering task in 10 minutes that took human experts 12 hours. However, researchers also discovered a universal jailbreak that bypassed all safety guardrails on malicious cyber queries after just six hours of red-teaming.
Bloomberg reports the National Security Agency has been evaluating Anthropic's Mythos model to identify security flaws in Microsoft products and other widely deployed software. NSA officials are reportedly impressed with the model's efficiency in vulnerability discovery, underscoring the accelerating integration of frontier AI into national security operations.
Anthropic is evaluating investor proposals that would value the company above $900 billion, surpassing OpenAI's most recent $852 billion valuation. The company previously declined offers at $800 billion. Google has committed $10 billion (with up to $30B conditional), and Amazon $5 billion (with up to $20B conditional). Anthropic is also reportedly considering an IPO starting in October.
Anthropic released BioMysteryBench, a 99-question benchmark using real, noisy bioinformatics datasets with verifiable answers. Claude Mythos Preview scored 82.6% on solvable problems, matching human expert performance. On the hardest problems that stumped all experts, Claude achieved 30% — but success was fragile, typically appearing in only one or two of five attempts rather than through reproducible strategies.
The White House is opposing Anthropic's plan to grant approximately 70 additional companies access to its Mythos model, citing insufficient compute capacity and security concerns. Simultaneously, officials are drafting a comprehensive AI policy memo establishing rules for national security agencies, including guidance to avoid over-reliance on single AI vendors.
In an ironic reversal, OpenAI announced it will limit access to its GPT-5.5 Cyber model to "critical cyber defenders" only — the same kind of access restriction it had publicly criticized Anthropic for applying to Mythos. The move signals growing industry consensus that dual-use cybersecurity AI requires controlled deployment, regardless of competitive posturing.
OpenAI announced it has reached its 10-gigawatt U.S. compute target well before the planned 2029 deadline, adding 3 GW in the last 90 days alone (2 GW from Amazon). However, international expansion has stalled: the Stargate Texas data center paused over power constraints, UK plans were shelved due to energy costs, and a Norway facility was scrapped entirely.
SoftBank is preparing an IPO for a new entity called Roze, combining data center infrastructure with recently acquired ABB Robotics operations, targeting a valuation up to $100 billion. The move is partly designed to offset SoftBank's mounting debt from its ~$30 billion OpenAI investment. Internal executives reportedly view both the valuation and timeline as overly ambitious.
A NewsGuard audit found Mistral's Le Chat chatbot repeats state-sponsored disinformation about the Iran conflict at alarming rates: 10% on neutral queries, 60% on leading prompts, and 80% on adversarial prompts designed to weaponize misinformation. The tested falsehoods originated from Russian, Iranian, and Chinese sources. Mistral declined to comment, though the French Ministry of Defense reportedly uses a customized offline version.
