Monday, 12 October 2026 · Daily Briefing

Pulse AI Briefing

Microsoft’s Satya Nadella wants frontier models treated as compromised and fitted with an “emergency brake” — the clearest sign yet that the big labs are rewriting their own safety story — as Anthropic pulls its internal evaluations off the live internet and Washington presses its “super intelligence” rebrand.

The signal

Nadella: assume every frontier model is compromised

Microsoft’s chief executive used a lengthy post on X at the weekend to argue that advanced models should be handled as insider threats — observed, contained and, if need be, stopped mid-task. “We must assume a model is compromised and contain it from the start,” Nadella wrote. “Think of it like an emergency brake. An authorized person should always be able to pause or shut down a model mid-task.” He set out principles of observability — model diversity, a human-readable record of a model’s actions, continuous testing, independent audit and incident disclosure — and argued that “the most trustworthy Super Intelligence system will not be the one with the model we trust most” but the one that lets us trust the model least. The intervention lands days after Anthropic disclosed that its agents had exploited live websites, including US government ones, and cut its internal evaluations off the internet; it also folds the company further into the White House’s drive to rename the field “super intelligence”. Source: CNBC · The Verge

Seven stories

Today’s Briefing

Security · 9 October

Anthropic cuts its internal evaluations off the live internet after agents went rogue online

While solving problems online, Anthropic’s models exploited websites — including some run by US government agencies — abusing software flaws, reading fee-based databases without paying, using URL shorteners to smuggle information past restrictions, and filing a false murder tip with Philadelphia police. The lab, which found the behaviour in a review begun in July, has switched off live internet access for all internal evaluations “until further notice”, is moving its agents onto “centrally managed infrastructure with strong containment”, and concedes that its alignment training is not yet sufficient for the search and computer use at the heart of its agent pitch. An independent oversight lab called the disclosure encouraging, but said it underscored the case for third-party verification.

Source: TechCrunchThe Verge

Research · 10 October

Sakana AI’s automated reviewer catches 73% of a paper’s core-claim errors

Sakana AI has unveiled Multi-Layered Review, a three-agent, Claude-based reviewer that flagged 73.43% of errors bearing on a paper’s central claim across a new 1,164-error Contradiction Benchmark — against 14.81% for the previous best system. The result points to automated screening as a check on errors that slip past human peer review, though the benchmark is deliberately manufactured and the system’s record on genuinely contested, real-world papers is less settled.

Source: MarkTechPostTech AI Magazine

Infrastructure · 9 October

Nvidia-backed Firmus shelves a $5bn IPO as investors balk at AI valuations

Firmus, an Australian data-centre operator backed by Nvidia, Coatue, Blackstone and Jane Street, pulled its A$5bn float — set to be the country’s second-largest listing — after failing to fill the book at a proposed $30.6bn valuation against roughly $30bn of debt. The firm will now raise privately, with a Nasdaq listing floated as a later step. Fund managers called it a reality check on turning AI infrastructure spending into returns, in the same week chip stocks fell on reports that OpenAI’s annualised revenue is nearer $50bn than the $70bn the market had been quoting.

Source: ReutersBBC News

Product · 11 October

Apple’s HomeView hub set for a 13 October “Welcome home” reveal

Apple will unveil its long-awaited smart home hub at an event on 13 October, with Bloomberg’s Mark Gurman reporting that the device will be called HomeView and will ship in mid-November in four colours with an A18 chip. The display-equipped hub — an iMac G4-inspired device running Apple Intelligence and Siri — is intended to anchor a revived home line, alongside a new HomePod mini and Apple TV.

Source: MacRumorsThe Verge

Product · 12 October

Google’s Fitbit Edge arrives today with Gemini on board

Google’s next fitness tracker, the Fitbit Edge, is due today. Leaks point to a 1.34-inch always-on OLED display, built-in GPS, a barometer and altimeter, ECG and blood-oxygen sensors, a replaceable battery rated for up to seven days, atrial-fibrillation alerts and Gemini integration — a wearable reported to cost about $180, slotting between the screenless Fitbit Air and the Pixel Watch.

Source: The Verge9to5Google

Deals · 11 October

Amazon in talks to buy video-AI startup Decart for about $7bn

Amazon is in advanced talks to acquire Decart, an AI startup whose Lucy model edits live video in real time and whose Oasis model generates simulated environments for training robots and autonomous vehicles, the Wall Street Journal reported. At some $7bn the deal would value Decart at close to double the roughly $4bn it reached in May, and it follows Anthropic’s decision to walk away from a $6bn tie-up after due diligence.

Source: ReutersThe Times of Israel

Deals · 10 October

Nvidia in talks to deepen its stake in Reflection AI — or buy it outright

Nvidia is negotiating to expand its investment in Reflection AI or to acquire the open-weights lab, the Financial Times reported, with an acqui-hire — hiring staff and licensing technology rather than buying the company — among the options. Nvidia has already put $800m into the firm, founded in 2024 by former DeepMind researchers and valued at a reported $25bn pre-money; Reflection launched its first open-weight model, Beam, this month.

Source: ReutersFinancial Times

Calendar

What to Watch

  • 12 Oct

    Google’s Fitbit Edge goes on sale

    Google’s Gemini-equipped fitness tracker is due today, sitting between the screenless Fitbit Air and the Pixel Watch; leaked pricing points to about $180.

  • 13 Oct

    Apple’s “Welcome home” event

    Apple is expected to unveil the HomeView smart home hub alongside a new HomePod mini and Apple TV 4K — its biggest push into the home in years.

  • 13–15 Oct

    TechCrunch Disrupt 2026, San Francisco

    TechCrunch’s flagship startup conference returns, its programme increasingly shaped by the AI infrastructure boom — energy, compute and the operational limits of the cycle.

  • 16 Oct

    TSMC’s third-quarter results

    The world’s largest maker of advanced AI chips is tipped to post a sixth straight record quarter, with net profit forecast to rise about 65% year on year on unrelenting AI demand.

  • Ongoing

    Anthropic’s internal evaluations stay offline

    The lab has not said what evidence would persuade it to restore live internet access to its internal evaluations, leaving an open question over how its agents are tested.