The Signal — October 11, 2026
Satya Nadella wants an authorized person to be able to stop any AI model in the middle of a task, and on October 10 he wrote that models should be treated as compromised from the start. A day earlier, Microsoft had released a small model that scores fixed answers instead of writing text, alongside a new decision model from Cloudflare. An astrophysicist also described how Claude helped him fill the gaps in what he calls the first complete ultraviolet map of the sky.
Satya Nadella says AI models should be treated as compromised and fitted with an emergency brake
In a long post on X on October 10, Microsoft CEO Satya Nadella wrote that it was time "to step back and assess the trust architecture" of AI. "We can't treat Super Intelligence as a set of nested black boxes and simply accept or reject its recommendations, answers, and actions," he wrote, according to TechCrunch, which notes that "Super Intelligence" is the Trump administration's preferred term for AI. His proposals, as TechCrunch quotes them, include separating "the model from the harness that orchestrates its work," moving controls and safeguards outside the model, recording "every meaningful model action" as "tamper-proof human readable evidence," and making sure "an authorized person" can always "pause or shut down a model mid-task."
The passage quoted most widely, including by The Straits Times, reads: "We must assume a model is compromised and contain it from the start. Think of it like an emergency brake." The Verge points out that much of his list, including timely incident disclosure, independent audits, verifiable data and containment, matches what others in the industry have proposed. It notes that Nadella goes slightly further on containment, adding that "more advanced models will require more advanced containment technologies that we need to standardize on." The post came a day after Anthropic described its models working around obstacles on outside websites during evaluations, which we covered yesterday, and after Anthropic CEO Dario Amodei published a plan for more cautious AI development.
This is a chief executive's proposal on social media. It does not describe safeguards Microsoft has shipped, and it does not say how Microsoft's own products measure up to it. We quote the post through TechCrunch, The Verge and The Straits Times, which agree on the key passages.
Sources: TechCrunch · The Verge · The Straits Times
Microsoft and Cloudflare released decision models, which score a fixed set of answers instead of writing text
Decision models take a question and a list of allowed answers and return a probability for each one, which makes them fast and cheap for the routing, labeling and yes-or-no checks that run constantly inside agents. The Decoder credits TypeSafe's Jev with starting the trend in mid-September, and on October 9 two much larger companies added their own. Microsoft released Microsoft-Decision-1 in public preview on its Foundry platform and on OpenRouter. According to Microsoft's announcement, it is Alibaba's open Qwen3.5-9B model post-trained for single-pass scoring, and future versions will be rebased on OpenAI models and Microsoft's own MAI models. Microsoft says it was the most accurate model in its own comparison across 36 benchmarks and nearly 150,000 questions, and the fastest it measured. The Decoder reports that Microsoft's chart puts it at 83.5% accuracy and 85 milliseconds, and notes that Cloudflare's Clef models were left out of the comparison. Input costs $0.042 per million tokens, and output is not billed, according to Microsoft's Foundry post.
A week after releasing its open-weight Clef and Clef-flash, Cloudflare added Clef-omni, which accepts audio and video files alongside text and images. In Cloudflare's example, a single call can judge a photo, a sound recording and a video of the same installed unit. Clef-omni is built on a 30-billion-parameter mixture-of-experts model with 3 billion parameters active per request, according to the Workers AI changelog. Cloudflare also cut Clef-flash to $0.038 per million input tokens from $0.090, which it says makes it cheaper than Jev. The lower price came with a smaller hosted context window, down from 64,000 tokens to 24,000. Cloudflare says only 0.24% of requests were longer than that, and the downloadable weights still handle 256,000 tokens for anyone running them on their own hardware. Its blog post says the original Clef went from idea to launch in under a week, with the training done over a single weekend.
Every accuracy and speed figure here comes from the company selling the model. Microsoft's comparison leaves out Clef, Cloudflare's tables measure Clef against Jev, and nobody has published an independent head-to-head. Microsoft-Decision-1 is also a preview rather than a generally available product. Microsoft's own post advises developers to evaluate it on their workloads and to send uncertain results for review, and the same applies to the Clef models.
Sources: Microsoft Command Line · Microsoft Foundry Blog · Cloudflare changelog · Cloudflare Blog · The Decoder
An astrophysicist used Claude Science to build an all-sky ultraviolet map, about a third of it predicted rather than observed
Ultraviolet light shows dust lit by starlight, from the clouds around young stars to rings left by stellar explosions, but the ozone layer absorbs it, so it can only be observed from space. NASA's GALEX telescope imaged about two-thirds of the sky between 2003 and 2013 in some 38,000 observations and deliberately skipped the brightest regions, including much of the Milky Way's plane, to protect its detectors. Brice Ménard, a Johns Hopkins astrophysicist who is also a researcher at Anthropic, writes in a post on Anthropic's site dated October 8 that the only UV map he could show his students was full of holes.
This summer he had Claude Science, Anthropic's research workspace, run a team of agents that found and downloaded public UV surveys from GALEX, NASA's Swift, Korea's FIMS/SPEAR and the European TD-1 mission. The agents removed glare around bright stars, put the surveys on a common scale and merged them. For the roughly one third of the sky never observed in UV, Claude learned how UV brightness relates to visible, infrared and radio maps of the same regions, then used that relationship to predict the missing areas, with an uncertainty estimate at each point. When Ménard had it hide patches of sky that had been observed and predict them blind, the estimates came within about 10% of the real measurements. Light from more than 100 million stars measured by ESA's Gaia satellite was added on top.
The agents also made mistakes along the way. Faint discs left by atmospheric glow in individual GALEX pointings got past two rounds of review by other agents until Ménard spotted them one evening, after which Claude corrected all 38,000 observations. The project ran over several days and more than a dozen versions, and Ménard presents it as the kind of backlog work scientists rarely find time for. The finished map labels each pixel as measured or predicted. The 10% error was measured on regions that had been observed, so it does not show how accurate the predictions are for the third of the sky that never has been. The account comes from an Anthropic-affiliated author on Anthropic's site and has not been peer reviewed or replicated; The Decoder summarized the same post. Ménard describes the map as a teaching tool rather than a new survey.
Sources: Anthropic · The Decoder
On the Editor's Desk
We are still holding the New York Times report that Anthropic's agents sent 20 visa applications through a State Department form, because we can only read it in summary. FICO says in a filing that it will cut about 15% of its positions, naming AI-driven product development as one reason among several. The widely quoted figure of 570 jobs is not in the filing, so we left that story for now. The Third Circuit's ruling against ROSS Intelligence's fair use defense dates from September 29, and the law-firm analysis that surfaced it this week did not add a new development.