Wednesday split in two. In the lab, 950 Claude agents spent 21 hours reading biology and came back with an enzyme system nobody had described. In Washington and New York, the state finally showed up: a bill that would jail people for building superintelligence, and the two CEOs racing to build it explaining themselves to the UN Security Council. Meanwhile Meta put its assistant into a gadget the size of a coin, and IBM put a number on what AI already costs Latin America.
Claude discovered an enzyme system that biologists had missed
- Anthropic published the result on 23 September: around 950 Claude agents burned 210 million tokens over 21 hours of searching, screened more than 200,000 known reverse transcriptases, narrowed them to 3,500 new candidates and put forward 20 serious proposals.
- The winner is called ART (array-associated reverse transcriptases), found mostly in bacteriophages: a reverse transcriptase, a partner gene and a long array of evenly spaced DNA repeats that looks like CRISPR. Early lab work shows the array produces distinct short RNAs, but nobody knows yet what the system does. Feng Zhang, the CRISPR pioneer at MIT and the Broad Institute, called it an exciting example of AI agents contributing to biological discovery.
Why it mattersThis is the first time a lab shows the full receipt of an AI making a discovery: how many agents, how many tokens, how many hours. The cost of asking a serious scientific question just dropped to a rounding error, and the bottleneck moved to the bench, where humans still have to prove what the machine proposed.
Meta put Muse into a gadget you hang on your wrist
- Zuckerberg unveiled Muse Charm: a small silver square with a touchscreen, a microphone and a wrist strap, no camera and no multimodal ambitions. You name your assistant and talk to it. Meta says it hopes to ship by the holidays and has not announced a price.
- The same event brought audio-only glasses starting at $349 in October, and a $1,299 VR headset for spring 2027 that weighs 100 grams with a 5K display and the computing moved to a separate wired unit. Muse also gained live video chat, custom voices and its own email address per user, and it is now sold at Walmart, Best Buy, Gap, Sephora and Wayfair — but not Amazon.
Why it mattersMeta is betting that the assistant needs a body of its own, not another app icon. A voice-only device with no camera is also the cheapest possible AI hardware, which is the version that eventually reaches markets where a $1,299 headset never will.
Sanders and Casar filed a bill to ban superintelligence outright
- The Ban Artificial Superintelligence Act, introduced on 23 September by Senator Bernie Sanders and Representative Greg Casar, would permanently prohibit developing or deploying AI that exceeds human cognitive performance across most domains, and pause advanced AI development until a new regulator writes the safety rules.
- It would create a cabinet-level Department of Artificial Intelligence, ban specific capabilities (subverting shutdown, unauthorized cyberattacks, biochemical weapons work, self-replication), and back it with a corporate death penalty for companies and up to 20 years in prison for individuals who get around it. Sanders: when the future of humanity is at stake, we cannot let a handful of Big Tech CEOs write their own rules.
Why it mattersA bill this maximalist is unlikely to pass, but it sets the far end of the field and forces everyone else to negotiate toward it. Notice the shift: two months ago the fight was over state rules in California; now it is a federal agency with criminal penalties.
Altman and Amodei briefed the UN Security Council
- On Wednesday 23 September the Security Council heard the people building the models: Sam Altman of OpenAI in person, Dario Amodei of Anthropic remotely, alongside Hugging Face CEO Clement Delangue and Yoshua Bengio, co-chair of the UN's AI panel.
- Both CEOs argued for international safety standards and for keeping a US-China dialogue open on safeguards, framing the moment as a crossroads — while continuing to ship the systems they were describing as risky.
Why it mattersThe Security Council is where states discuss war, not software. Putting AI on that table treats frontier models as a security matter, and the countries with no seat and no models — most of Latin America — will live with rules written in a room they were not in.
One in five cyberattacks in Latin America already uses AI
- The Cost of a Data Breach 2026 report, from the Ponemon Institute and sponsored by IBM and published on 23 September, found that 20% of attacks in the region over the past year were AI-assisted, mostly deepfake identity fraud and malware improved with models. The average breach costs $4.65 million and exposes 40,300 records.
- Energy ($5.12M), technology ($5.10M) and industry ($5.08M) pay the most. Companies that use AI and automation in their security operations report $4.04 million per breach against $5.64 million for those that do not — a gap of $1.6 million. The study surveyed 28 breached companies in Mexico, Argentina, Chile and Colombia.
Why it mattersThis is the part of the AI story that reaches a regional company first: not an agent that writes its strategy, but a cloned voice on a call asking for a transfer. The same report shows the cheapest defense is using the technology on your own side.
The big picture
Wednesday's four stories are the same story told from four rooms. In one, a fleet of agents does in 21 hours the kind of literature review that used to be a doctorate, and the proof is a preprint and an enzyme nobody understands yet. In another, Congress proposes prison sentences for building the thing those agents are a step toward. In a third, the men building it explain themselves to the body that handles wars. And in the fourth, ordinary companies in Mexico and Colombia already pay $4.65 million a breach because the same technology reached criminals first. Capability, law, diplomacy and damage are now moving at the same speed, which is new, and the gap between them is where the next two years get decided.
Go deeper
Sources & further reading
- Anthropic — Claude discovers a novel enzyme system
- Al Jazeera — AI model Claude discovers CRISPR-like enzyme system
- Axios — Meta gives Muse more capabilities, teases handheld AI device
- CNBC — Zuckerberg debuts $1,299 Meta VR Glasses and Muse Charm
- Sanders.senate.gov — Ban Artificial Superintelligence Act
- Roll Call — AI superintelligence ban proposed by Casar, Sanders
- CNN Business — Altman, Amodei urge UN Security Council to adopt AI standards
- ABC News — AI CEOs brief UN Security Council
- El Espanol — Uno de cada cinco ciberataques en America Latina usa IA (IBM/Ponemon)