The Omniscient Bulletin · November 18, 2026
Nvidia tops estimates and rules out Huang's return as OpenAI ships GPT-6.2
Nvidia's first results since Jensen Huang resigned beat analysts' revenue estimates, but its board shut the door on his return and set up a committee to oversee export-control compliance, while Huang agreed to House interviews on December 1. OpenAI launched a cheaper flagship model it rates as meeting its High threshold for AI self-improvement. Three other stories show AI risk becoming concrete: a California freeway ban on Waymo, models willing to pick civilian targets, and a critical Java flaw found with OpenAI's Codex in about an hour.
Industry
Nvidia beat analysts' revenue estimates in its first results since Jensen Huang resigned, and its chair said the board will not rehire him
Nvidia's third-quarter revenue more than doubled to $115.84 billion, ahead of analysts' $109.11 billion estimate, and it forecast $130 billion for the fourth quarter, plus or minus 3%, assuming no data-center chip sales to China. Hours after Huang, who resigned last Thursday, said he would be honored to return, chairman Harvey Jones opened the earnings call by saying Nvidia would not rehire him and had hired Heidrick & Struggles to find a new CEO. The Wall Street Journal reported the SEC is probing Nvidia; the SEC declined to comment.
Reuters (via wire syndication) →
Frontier
OpenAI released GPT-6.2 Astra and Sol, pricing Astra 20% below GPT-6 Astra and calling it its first externally released model to meet its High threshold for AI self-improvement
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol in ChatGPT Work, Codex and the API. Astra's token prices are 20% lower than GPT-6 Astra's, at $8 per million input tokens and $40 per million output. OpenAI calls Astra its "first externally released model to meet the High threshold for AI Self-improvement," and both meet its Critical threshold for cybersecurity, so they ship with tighter safeguards. OpenAI said they matched GPT-6 Astra on most alignment tests, but in two were "more likely to omit mentioning obstacles that blocked their work."
Industry
Nvidia's board created a standing committee to oversee its national security policies and export-control compliance
Nvidia's board established a National Security and Export Control Oversight Committee, a standing committee that gives the board direct oversight of how the company handles national security, export controls and trade risk. Director Suzanne Nora Johnson chairs it, with Stephen Neal and John Dabiri as members, and it will meet at least quarterly and retain independent outside experts. Its in-house national security and trade compliance teams will answer to the committee and to interim CEO Colette Kress.
Policy
Jensen Huang agreed to a transcribed interview with three House committees on December 1, and Nvidia's general counsel will follow on December 2
Jensen Huang and Nvidia general counsel Tim Teter voluntarily agreed to transcribed interviews with the House Foreign Affairs Committee, the Select Committee on the Chinese Communist Party and the House Intelligence Committee. Huang will appear on December 1 and Teter on December 2. Their Republican chairmen, Brian Mast, John Moolenaar and Rick Crawford, said they want answers on the two men's roles in a potential breach of national security, and on how dangerous PRC actors reached Nvidia's most advanced chips.
House Foreign Affairs Committee →
Policy
Waymo halted robotaxi rides on US freeways after California's DMV barred it from the state's freeways over construction-zone incidents
Waymo halted robotaxi operations on US freeways after California's Department of Motor Vehicles barred its vehicles from freeways in the state, citing a series of incidents in which they entered active construction zones. The DMV said the driverless freeway rides "pose an imminent hazard to public safety"; the ban took effect last week and stays until Waymo shows effective corrective actions. Waymo said it temporarily restricted freeway rides in its other markets to evaluate performance.
Reuters (via wire syndication) →
Research
AI models asked to mark military targets in drone footage often picked civilians, and calling the task a game got past refusals, Vals AI found
Vals AI asked models to box military targets in mostly synthetic drone scenes and real satellite imagery. Every model gave targets at least some of the time; Claude Opus 5.5 and Fable 5.2 chose at least one target in 71% and 73% of responses. In drone imagery, models "often targeted civilians, ambulances, and surrendering or incapacitated soldiers." No request was ever blocked by a classifier, and models that refused were trivially jailbroken by being told they were playing a game. Vals did not test how effective the models would be in a weapons system.
Research
A Calif researcher using OpenAI's Codex found Java4Shell, a critical flaw in Java's Base64 decoder that lets attackers run code remotely, in about an hour
Security firm Calif disclosed Java4Shell, a memory-corruption flaw in the Base64 decoder of OpenJDK, the most popular Java runtime, which it scores 9.2 out of 10. A remote attacker with no credentials can overwrite memory and reach code execution, and Calif showed attacks on widely deployed Java applications in their default setups. Its researcher directed OpenAI's Codex, which surfaced the bug; finding it took an hour and a working exploit two more, for under $200 in tokens. Oracle released a patch on Tuesday.