The Omniscient Bulletin · 2026-09-30
The Omniscient Bulletin — September 30, 2026
The day Washington asked the AI industry to police itself, and the industry's own safety reports showed what it is policing.
President Trump released a voluntary accord, signed by six AI industry leaders, that pledges internal controls, an outside auditor and board oversight for frontier models. OpenAI rated its new GPT-6.1 Sol Critical in cybersecurity, and Anthropic said in its own tests that an open-weight Chinese model now builds exploits almost as often as Claude Mythos Preview. Also: the first US appeals ruling on AI training and copyright, OpenAI's always-on dots agents, Anthropic's largely non-cancelable compute bill, and OpenAI's 1.4 trillion dollar funding talks.
Policy
Trump released a voluntary accord, signed by six AI industry leaders, pledging internal controls, an outside auditor and board oversight for frontier models
After a White House lunch on Tuesday, President Trump released the "White House Accord on Super Intelligence," signed by Sundar Pichai, Dario Amodei, Mark Zuckerberg, Elon Musk, Jensen Huang and OpenAI president Greg Brockman. It says every company training frontier models should monitor capabilities and alignment during training and deployment, empower an internal team to check those controls, partner with an independent external auditor, and have an independent board committee oversee it all. It is voluntary, though Trump called it "morally binding."
Washington Examiner (full text of accord)→
Frontier
OpenAI released GPT-6.1 Sol and rated it Critical in cybersecurity, so it ships with the same safeguards stack as GPT-6 Astra
OpenAI launched GPT-6.1 Sol, which it calls comparable in capability to GPT-6 Astra while stressing its speed and affordability. It rates Sol Critical in cybersecurity, High in biological and chemical capability and below High in AI self-improvement, with Astra's safeguards applied. On its own internal test, built from recently disclosed vulnerabilities to limit contamination, Sol reached arbitrary code execution 21.5 percent of the time, against 31.5 percent for Astra and 5.5 percent for GPT-6 Sol.
Frontier
Anthropic said Zhipu's open-weight GLM-5.3 builds end-to-end cyber exploits almost as often as Claude Mythos Preview, with safeguards that are easily bypassed
In its own testing, Anthropic said Zhipu AI's GLM-5.3 built working end-to-end exploits in 50 of 410 attempts, against 56 for Claude Mythos Preview, and shipped with no meaningful protections against misuse. Anthropic's researchers stripped GLM-5.3's refusals in about 2,200 GPU hours, roughly 4,400 dollars of compute. Anthropic concluded that "a critical threshold in freely accessible capabilities has now been crossed" and said governments should safety-test sufficiently capable models, including GLM-5.3's successors.
Policy
A federal appeals court upheld Thomson Reuters' copyright win over Ross Intelligence, the first US appeals ruling on AI training and copyright
The Third Circuit on Tuesday upheld a ruling that Ross Intelligence could not claim fair use for copying Westlaw headnotes to train its AI legal search engine. No other copyright fight over AI training has reached a US appeals court, though this one concerns a search tool, not generative AI. The court's reasoning is sealed for now. Ross closed its platform in 2021, blaming the expense of the suit. Reuters, which reported the decision, is owned by Thomson Reuters, the plaintiff.
Agents
OpenAI unveiled dots, always-on agents powered by GPT-6 Astra, though its live demos repeatedly failed on stage
At its DevDay, OpenAI launched dots, always-on agents that pursue a user's goals across apps on their own and can be messaged on Slack and Teams. CEO Sam Altman said dots had been safety tested and run on GPT-6 Astra, and OpenAI highlighted custom rules users can set on what the agents may do. On stage, the agents failed several times to respond with spoken updates. OpenAI also said more than 35 million people a week now use Codex and ChatGPT Work.
Industry
About 80 percent of Anthropic's 518 billion dollar compute plan is non-cancelable or owed regardless of use, while up to 84.5 billion dollars with xAI is largely cancelable on 90 days' notice
Anthropic's confidential IPO prospectus, seen by Reuters, sorts its 518 billion dollar buildout by how easily it can be canceled: about 80 percent is non-cancelable or payable regardless of usage, including at least 111.1 billion dollars with Google, 110 billion with Amazon, 31.4 billion with Microsoft and about 161.2 billion in Broadcom-related equipment leases. Up to 84.5 billion dollars of Nvidia-based xAI capacity through 2029 is largely cancelable on 90 days' notice. Editor's note: September 29th's issue reported the 518 billion dollar total; this adds the cancelability breakdown.
Industry
OpenAI is seeking at least 30 billion dollars at a 1.4 trillion dollar valuation as a bridge in place of an IPO, as its revenue run rate nears 70 billion dollars
OpenAI aims to raise at least 30 billion dollars at a valuation of about 1.4 trillion dollars before the new money, Bloomberg reported, in a round meant to serve as bridge financing in place of an IPO. The discussions are early, and the terms may still shift. OpenAI closed a round in March with 122 billion dollars committed at an 852 billion dollar valuation. Separately, its annualized revenue run rate is nearing 70 billion dollars, up more than 70 percent since the start of the third quarter.