The Omniscient Bulletin · 2026-08-20
The Omniscient Bulletin — August 20, 2026
OpenAI says it can police its models without reading your prompts. Five agencies say AI-written attack code is already inside the controllers at water plants. Google buys the right to buy Marvell. And Senate Republicans warn that data centers may cost them Ohio.
Yesterday OpenAI put a price on watching its own models. Today it argues the watching need not come with reading, and aims that at a rival now demanding logs. Hours earlier five federal agencies described what the watching is for, and it was not hypothetical: AI-written exploit scripts, dressed as monitoring tools, already probing the controllers that run water and power. The money and the politics both moved in the same direction.
Frontier
OpenAI says it can catch misuse without its staff reading your prompts, while Anthropic's top models still require 30 day logs
OpenAI is previewing Private Safety Processing, which it says spots misuse patterns across related interactions without giving OpenAI staff access to the underlying content. Customer data stays on infrastructure the customer controls, or sits encrypted under keys the customer holds; when something trips, OpenAI gets only a narrow signal naming the type of activity. The post notes that some recent frontier deployments require customers to let their provider retain sensitive content. Anthropic requires 30 day retention on Fable 5 and Mythos 5.
Policy
Five federal agencies say attackers are already using AI-written code to reach the controllers at water and power plants
The NSA, CISA, FBI, the Energy Department and the EPA issued a joint advisory on Siemens S7 programmable logic controllers, saying attackers are deploying AI-generated Python scripts built on the open source snap7 libraries and dressed as legitimate monitoring tools. The agencies say the actors scan the internet for exposed controllers left on default or minimal authentication, and name water, energy, chemical and critical manufacturing among the sectors most targeted. In its own words, this is not a theoretical risk, it is an active threat.
Frontier
OpenAI vetted security researchers for a restricted cyber tier, then locked out five who all live outside the US and Europe
Five researchers told TechCrunch that their access to Trusted Access for Cyber, OpenAI's vetted program offering advanced models with fewer cybersecurity guardrails, stopped working on Wednesday, showing messages that their identity could not be verified or that the account was ineligible. OpenAI blamed a technical issue affecting a limited number of users, said it was an issue on our end, and asked them to re-verify. Every researcher the outlet spoke to lives outside the US and Europe. The tier launched on August 10.
Compute
Google can buy $12.2 billion of Marvell stock, but only if it buys $120 billion of Marvell's chips first
A Marvell filing discloses a warrant issued to Google on August 18 for up to 58,970,907 shares at $206.58, attached to a July 29 agreement covering custom silicon that plugs into the TPU ecosystem: inference accelerators, storage and network controllers, memory interface controllers and near memory compute. Only 1,360,867 shares vest on a clock. The rest vest in 240 equal tranches, one for every $500 million of custom product revenue, through fiscal 2033. Fully earning it means Google buying $120 billion of Marvell silicon.
Agents
Wired rebuilt an unreleased police AI from the code Flock's own login page was handing out to anyone who asked
Wired reconstructed an unreleased Flock Safety tool from a cache of more than 450 files the company's own login portals served to anyone who loaded them. Originally called Nightshift and since renamed OS Investigate, it ships with 69 prewritten prompts officers can edit and submit. One reads: "Find me witnesses based on vehicles most seen in [neighborhood] during [last 14 days]." The code describes 45 tools reaching plate scans, arrest records, case files, 911 dispatch logs, ballistics and commercial databases holding Social Security numbers and relatives.
Policy
Senate Republicans told AI companies privately that data centers could cost them Ohio, and that every other state is watching
In a memo obtained by Axios and headlined Ohio Data Center Risk, the National Republican Senatorial Committee tells top AI companies that Democrats have made data centers a centerpiece of the campaign against Senator Jon Husted, and that it is working. If he loses and data centers get the blame, the memo says, politicians across the country will take notice, and they will not go near the next one. Separately, AI executives and Republicans told Axios that internal polls put data centers about as popular as spent nuclear waste.
Industry
Stripe told investors the singularity began on January 1, and that it intends to stay private all the way through it
In a letter obtained by Axios, Stripe said January 1 marked the beginning of the singularity, which it calls a major inflection point in long term trends. It also told investors that being privately held is the best structure for such a consequential moment, which Axios reads as putting any listing on indefinite hold. First half revenue rose 41% and free cash flow 43%. Stripe says 88% of the Forbes AI 50 build on it, and that revenue from AI and crypto companies has more than doubled.