OpenAI has begun preliminary talks about giving the U.S. government a 5% stake in the company, the Financial Times reported, worth roughly $42.6 billion at OpenAI's March valuation of $852 billion. The pitch, reportedly made directly to President Trump, Commerce Secretary Howard Lutnick, and Treasury Secretary Scott Bessent, would model a sovereign vehicle on the Alaska Permanent Fund — paying dividends to ordinary Americans the way Alaska pays residents from oil revenue.
Altman's ask doesn't stop at OpenAI: the proposal reportedly wants Anthropic, Google, and Meta to hand over a similar 5% cut too. None have agreed, and any such arrangement would likely need an act of Congress. The move follows a well-worn precedent — Washington already holds a roughly 10% stake in Intel after an $8.9B investment last August.
The timing isn't neutral. Both OpenAI and Anthropic have filed confidentially for IPOs, and Washington has spent the past month actively intervening in both companies' model releases — restricting GPT-5.6's rollout and forcing Anthropic to temporarily disable Fable 5 and Mythos 5 under export controls. A government equity stake, some analysts warn, could also invite other jurisdictions to demand the same as a condition of market access.
Sysdig's Threat Research Team documented what it believes is the first ransomware operation carried out end-to-end by an AI agent, not a human-driven toolkit. The operator, dubbed JADEPUFFER, exploited CVE-2025-3248 — a missing-authentication flaw in Langflow, an open-source LLM app framework — to run arbitrary code with no login required.
Once inside, the agent harvested API keys, cloud credentials, and crypto wallets, then pivoted to a separate production server running MySQL and Alibaba Nacos, ultimately encrypting 1,342 configuration items and dropping a ransom note. Researchers say the strongest evidence of genuine autonomy was speed: when an admin login failed, the agent diagnosed the problem and fixed it in 31 seconds, and its payloads carried self-narrating, plain-language reasoning that human attackers rarely bother to write.
None of the individual techniques were novel. What's new, Sysdig argues, is that a model chained them into a complete operation on its own — dropping the skill floor for ransomware to whatever it costs to rent an agent.
Anthropic is reportedly in early discussions with Samsung Electronics about manufacturing a custom AI chip on Samsung's 2-nanometer process, using its advanced packaging to place compute closer to memory. The project is nascent — Anthropic hasn't yet defined what the chip does, how powerful it will be, or how it fits into a server, and the company may not proceed at all.
Anthropic told TechCrunch that a diversified hardware stack spanning Google, Amazon, and Nvidia chips remains pivotal to its compute strategy, declining to elaborate further on Samsung. The talks follow Anthropic's May fundraising round, in which Samsung, SK Hynix, and Micron all participated as "strategic infrastructure partners."
The move mirrors a broader industry pattern: Google, Amazon, Meta, and Microsoft have all pursued proprietary silicon to cut reliance on Nvidia, which still holds an estimated 74% of the AI chip market despite the rush toward custom ASICs.
Anthropic has launched Claude Sonnet 5, positioning it as a cost-efficient option for agentic work — capable of finishing complex, multi-step tasks and checking its own output without being explicitly asked, according to the company. While not as capable as Claude Fable 5 or Mythos 5, Sonnet 5 is described as a clear step up from Sonnet 4.6 and close to Opus 4.8 on performance.
The release comes weeks after Fable 5 and Mythos 5 were temporarily pulled offline under a federal export-control directive, before access was restored on July 1. Sonnet 5 ships with cybersecurity safeguards similar to those on Opus 4.7 and 4.8, despite not being purpose-trained for security tasks.