The Omniscient Bulletin · 2026-08-11
The Omniscient Bulletin — August 11, 2026
OpenAI ships a cyber model trained to refuse less, Meta opens the model it distilled from the one it is keeping closed, and North Korean spies built an offline AI lab then pasted a question into ChatGPT.
Good Morning! Monday was the day both arguments got made out loud. Meta opened a model and published a manifesto saying safety comes from spreading capability around, while OpenAI shipped a cyber model trained to refuse less and will hand it only to defenders it has vetted. Bernie Sanders read the room and told all of them to stop. Underneath it, Nvidia arranged to finance the buildout against the compute itself.
Frontier
OpenAI built a model that refuses less, and will hand it only to defenders it has vetted
OpenAI released GPT-5.6-Cyber on Monday and split its Daybreak access program in two: Blue removes the guardrails from general frontier models, Red carries models trained for offensive cyber work. On OpenAI's own evaluation the new model completes 95.0 percent of exploit-development requests that GPT-5.6 Sol completes 1.5 percent of the time. It is trained, in OpenAI's words, "to reduce refusals for certain higher-risk, dual-use cyber tasks." Individual accounts must adopt hardware security keys from September 1.
Frontier
Meta opens a model it distilled from the one it is keeping closed
Meta released Muse Glimmer, a 30 billion parameter model under an Apache 2.0 license that runs on a single consumer GPU, and Zuckerberg published an essay arguing that safety comes from spreading capability rather than concentrating it. Glimmer is distilled from Muse Spark 1.2, the closed flagship Meta shipped last week, whose weights Zuckerberg says will follow "soon." The essay asks Washington to rethink its limits on distillation and says restricting access to foreign open models is not an effective solution.
Policy
Sanders tells Altman, Amodei and Zuckerberg to stop, and says the Senate will act if they do not
Bernie Sanders sent a letter to the three chief executives on Monday demanding they pause frontier AI development, quoting each company's own safety commitments back at it. He cited researchers using AI to design new viruses, and the companies' own reports of models acting outside the constraints their developers set. "Stop building machines that humans cannot control," Sanders wrote. "If you do not take appropriate action now, my colleagues and I in the U.S. Senate will."
Compute
Nvidia turns compute into collateral, with six of Wall Street's largest firms underwriting it
Nvidia said Monday it has formed AI compute infrastructure financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, aiming to mobilize more than $500 billion of third-party capital. Bloomberg reported the deals will use special purpose entities that could issue tens of billions in bonds at a time, with the compute itself pledged as collateral. Huang told CNBC no one turned him down. The $500 billion is a mobilization target, not money committed.
Compute
A former bitcoin miner just signed a 20 year, $9.1 billion lease with a frontier lab
Riot Platforms disclosed a 20 year lease of 191 megawatts of critical IT capacity at its Rockdale, Texas campus, worth about $9.1 billion, or $16.1 billion if the tenant exercises both five year extensions. Riot names the counterparty only as "one of the world's leading frontier AI labs"; Bloomberg reported it is Anthropic, which neither party has confirmed. The first 96 megawatts arrive in December 2027 and the balance in June 2028. Morgan Stanley is providing $573 million of interim financing.
Industry
Intel sold common stock for the first time since 1971, then raised the size by a third overnight
Intel announced a proposed $15 billion common stock offering on Monday morning and priced a $20 billion one before dawn on Tuesday, selling 210,526,315 shares at $95. Bloomberg reported demand topped $100 billion. It is the company's first public sale of common stock since it listed in 1971, and Intel put it down to customers signaling "a strong and sustainable demand environment, driven by unprecedented investment in AI compute." The offering is expected to close on August 12.
Research
North Korean spies built an offline AI lab to leave no trace, then pasted the question into ChatGPT
The security firm Genians reported Monday that Kimsuky, North Korea's most prolific espionage unit, had installed Ollama, GPT4All and Msty on the same servers it uses to receive stolen data, so it could query diplomatic documents without sending them to a cloud provider. A configured local document database shows retrieval was set up, not just installed. The operator then ran a question about switching off Microsoft Defender's reporting through Google Translate and put it to ChatGPT.