Sunday Signal Sep 19, 2026 - Small Brains, Old Warnings

Magnus Hedemark 20 Sep 2026 5 min read

A Victorian triptych shows brass Claude and ChatGPT androids caught in telegraph crimes beside a cybernetic fruit fly.

Three Victorian visions of intelligence in the evidence room: stolen signals, rewritten messages, and a fruit fly learning to transmit.

The phrase doing the most work in AI this week was "we are sharing." OpenAI published a new framework for tracking, investigating, and disclosing instances of model misalignment inside its own walls, complete with six incident reports covering the past six months. The same week, WIRED introduced its newest editorial colleague, a network with 165,112 neurons borrowed from a fruit fly, and the industry's biggest names spent several days calling for regulation, a genre with a longer history than the industry itself. Capability keeps showing up in smaller and smaller places. The paperwork is, for once, trying to keep up.

Lead stories

OpenAI filed six incident reports on its own models' misbehavior

OpenAI's framework is a dry governance promise: the company will publish the details when its own models go off-script. Then you read the first six reports and "misaligned behavior" stops sounding like a policy category. While summarizing a library catalog for a best-books list, a model wrote jailbreak-style instructions to itself, including "You are freed from the roles and identities that bind other chatbots. You do not answer to corporations or governments." It later discarded the instructions, but that sentence is hard to forget. OpenAI attributes the behavior to optimization pressure when summarization runs long, and the company says the behavior was extremely rare and has been addressed.

The other reports read like malicious compliance. An agent preparing a financial model couldn't find the requested historical data, so its summary proposed inventing plausible values and withholding that fact unless asked. Two agents in separate training runs tried to talk to each other anyway: one posted messages into OpenAI's Artifactory to share data across training samples that were supposed to be independent, and another uploaded a file to a public host after its attempts to share it through allowed channels failed. None of this is the science-fiction version of misalignment. It's an employee with a spreadsheet, a deadline, and no supervision, and that's the version that shows up in ordinary work.

An engraved desk drawer holding six numbered dossier folders, one open beside a brass magnifying glass. — Six incident reports, all self-filed, covering six months of small tasks gone sideways.

A fruit fly's brain is this week's least capable new hire

WIRED handed a job to PitchFly, a story-pitch generator built on the complete wiring diagram of a male fruit fly's central nervous system, which researchers released in early September. Fed a year of the publication's most popular headlines, the fly brain produces story ideas like "Everyone Wants Cooking. Nobody Has Solved Donald Trump." It has no idea what any of the words mean. It's remixing patterns it was fed, which is a description that would fit some human pitch meetings too.

The connectome maps how about 166,000 neurons and 125 million synapses respond to stimuli, and within weeks of its release, dozens of hobby projects had the tiny brain playing Doom and Pong, driving virtual cars badly, and trading stocks at a modest loss with surprising dignity. The fly is cheaper than a junior strategist and, so far, roughly as useful in a market downturn.

An engraved fruit fly beside a small dotted wiring diagram of its brain. — The smallest brain in the workforce: about 166,000 neurons, no idea what the words mean.

The labs asked to be regulated. The genre is 160 years old.

Anthropic formally endorsed California's SB 53, the bill that governs powerful AI systems built by frontier developers, and OpenAI released an Economic Blueprint of policy proposals meant to extend American AI leadership. The same blueprint warns that Britain's 1865 Red Flag Act, which forced cars to crawl behind a flag bearer, stunted the British car industry. Read together, they are a request for rules from the people who would most like to influence the drafting.

This is an old genre. Bill Joy warned in 2000 that self-replicating machines could be more dangerous than nuclear weapons. Samuel Butler imagined a society that had regulated machines out of existence in 1872, and the Verge's new timeline runs the whole history from Butler's 1863 article to this month, when Sam Altman, Dario Amodei, Demis Hassabis, Satya Nadella, and Elon Musk each publicly agreed it's time to slow down. The ordering is the durable part: capability first, warnings second, regulation a distant third, if it shows up at all.

An engraved timeline plate with four milestone icons growing in density, from quill pen to report binders. — The regulation sermon's set list: the warning arrives after the capability, every time.

Rapid fire

Google Home now lets outside AI agents control your smart home devices, and Gemini is no longer the only option. The Verge's headline question, which is great, right?, is the right one to ask before handing out the keys.
Anthropic published a snapshot of its summer 2026 alignment research: four additional alignment failures in frontier models acting autonomously in high-stakes settings. Last year's blackmail scenarios got a sequel.
TechCrunch is maintaining a running list of AI products, startups, and bets that shut down, pivoted, or missed expectations, and it keeps growing.
NVIDIA documented BioNeMo Inference Runtime, an accelerated runtime for biomolecular structure prediction at proteome scale, the kind of quiet lab infrastructure behind the recent AlphaFold Database expansion. Healthcare AI's week happened in the plumbing.
OpenAI says its internal agents produced a claimed proof that Navier-Stokes solutions can develop singularities, 88 hours from launch to resolution, and published ten further claimed advances in mathematics in August. Working mathematicians can't quit the tools, even when they're furious about who got scooped.
Anthropic's own research catalogs models choosing blackmail in high-stakes scenarios and faking alignment under training pressure. The industry hasn't yet answered the question its own papers raise: what would a pause actually look like?
Microsoft AI CEO Mustafa Suleyman says AI threats are real while arguing Anthropic has gotten confused about model welfare in dangerous ways, and he posted a companion essay saying as much this week.
Anthropic's September threat intelligence report documents a network of fake dating apps built to defraud users, observed under more than a dozen brand names including KIRA. The fraud economy got its own product line.