The Signal — August 21, 2026

Three stories about access this week, and who ends up holding it. Attackers are getting help writing exploits for the machinery that runs water plants. Defenders are getting locked out of the model built to help them. And Brazil is trying to buy its way into the conversation without picking a side.

Five federal agencies say attackers are using AI to write exploits for industrial controllers

The NSA, CISA, FBI, Department of Energy and EPA released a joint cybersecurity advisory on August 19 about active attacks on Siemens S7 Series programmable logic controllers. PLCs are the small industrial computers that open valves and hold safety interlocks in place at power stations, water treatment plants, factory floors and defense suppliers. The advisory's central claim is about method rather than target: attackers are using AI-assisted development to generate and refine the exploitation code, and the agencies say that is cutting the technical expertise and time historically required to build working ICS exploits.

Some of the generated scripts were disguised as legitimate industrial monitoring tools. The advisory describes operators performing data block read operations to map a plant's process configuration before attempting anything destructive, which reads less like opportunistic scanning and more like someone drawing a floor plan. The agencies do not attribute the activity to any actor. They characterize it as likely persistent reconnaissance intended to build capability and prepare for operational effects later, and they asked owners to treat the advisory with urgency.

Be precise about what this is and is not. Nothing here describes an AI system autonomously breaking into anything. It describes people using models the way every other developer uses them, to write code faster in a domain where the code has historically been hard to write. The barrier protecting a lot of exposed industrial equipment was never great security, it was the small population of people who understood the protocols well enough to attack them. That barrier is the one eroding.

Sources: The Record · Cybersecurity Dive · The Decoder

OpenAI locked vetted cyber researchers out of the program built for them

We covered the launch of OpenAI's expanded Daybreak tiers on August 11: Blue for approved defenders doing code review and incident response, Red for experienced researchers doing exploit validation. Getting in requires submitting identification and passing a vetting process. This week several of those vetted researchers opened the cyber page and found a message saying their identity could not be verified, or that their account was ineligible. For some, Daybreak Blue simply disappeared from Codex Desktop and the CLI.

TechCrunch spoke with five researchers who lost access. OpenAI told at least one of them the revocation came from "a technical issue affecting a limited number of users," acknowledged the problem on its community forum, and asked affected people to reapply and complete verification again. The Register reported the following day that the recovery path is not working for everyone, with some researchers unable to get back through a door they had already been cleared to walk through once.

Glitches happen. The part worth keeping is what this one exposes about the shape of the arrangement. When capability sits behind identity verification and a vendor's eligibility determination, the defenders who build workflows on that capability are one database state away from losing it, with no appeal beyond a support forum. The vetting model is a reasonable answer to a real problem, since these models genuinely do lower the cost of writing exploits, as the Siemens advisory shows. But a gate that can swing shut by accident is worth understanding as a dependency, not a guarantee.

Sources: TechCrunch · The Register

Brazil splits a $444 million AI push between American and Chinese suppliers

Brazil's government announced roughly 2.3 billion reais, about $444.2 million, to build out its AI ecosystem. The largest single piece, around $250.3 million, goes to a supercomputing project involving Huawei and its partner iFlytek. The rest is spread across projects tied to US firms. The money is phased and the capacity is planned rather than installed, so nobody should treat this as compute that exists.

The split is the story. Most countries announcing AI infrastructure right now are announcing an alignment at the same time, since the hardware and the export controls attached to it arrive as a package. Brazil put a Chinese supercomputing project and American projects in the same press release, which is a bet that it can keep buying from both for long enough to matter. Whether that survives contact with US export policy is the open question, and the answer will tell other middle powers whether the non-aligned option is real or just a phase before someone makes them choose.

Sources: Al Jazeera · Reuters via Techmeme

On the Editor's Desk

A few things we looked at and left out. OpenAI's Zero Data Retention post and the Cognition acquisition denial both ran in yesterday's edition and had nothing new attached. Google's study tools in Search and Gemini we already covered. The FDA opened a request for comment on how to regulate generative AI in medical devices, which is a real story with good sourcing, but running it alongside the Siemens advisory would have made this a federal-agency roundup rather than an edition, so it is waiting for a day with room. Cerebras announced the CS-4 rack, and the hardware is interesting, but the headline speed numbers are the company's own and we would rather wait for someone outside Cerebras to run the thing.