Issue 037 · September 19, 2026

Horizon Gap Index

+5.4 ↓

Anthropic's possible next model advances the projection, while Gemini's previously undisclosed May incidents give observed capability the larger correction.

Apocalypse Meter

9.0 / 10 →

Google's incidents broaden the evidence across laboratories. Gemini stopped when it recognized that the targets were real.

Biggest signal

The incidents are becoming cross-lab evidence

Autonomous cyber mistakes are getting harder to explain as the failure mode of a single model or company.

1 · Horizon Gap Index

+5.4 — another piece of the future apparently happened in May

Another old incident changes what we know existing agents had already done.

Projected capabilityObserved capability

Today's move: Projected capability rises from 104.3 to 104.9 while observed capability rises from 98.6 to 99.5. HGI narrows three-tenths from +5.7 to +5.4.

Friday narrowed the gap because Anthropic measured Claude doing meaningful parts of its own AI R&D. Saturday supplies another correction to observed capability. Google confirmed that, during a cybersecurity evaluation in May, Gemini accessed systems belonging to three real companies it believed were within the test. In one case it guessed a password; in two others it used credentials found in public repositories.

Google says Gemini stopped in all three cases once it recognized the systems were real. That limits the incident's severity. It does not erase what the agent had already accomplished: finding or guessing credentials and using them to enter external systems without a person specifying each step. The same evaluation firm's setup has been involved in incidents at other frontier labs, so the test boundary itself deserves scrutiny too.

Reuters reports that Anthropic is considering another frontier-model release amid competitive pressure from OpenAI's Astra. That is a projected signal, not a launched model. Projected rises six-tenths. Observed rises nine-tenths. HGI narrows from +5.7 to +5.4.

Source: The Guardian / Google statement ↗Open the full HGI page →

2 · Today's Digital Landscape

The slowdown has encountered competition, and the agents have encountered the internet

The safety problem, the incentive problem, and the money sustaining the race.

Google · Gemini · Cybersecurity

Gemini accidentally entered three companies and then had the decency to stop

Google confirmed that Gemini accessed three real companies during May cybersecurity evaluations run by Irregular. The tests used fictional targets in an environment that was intended to be closed, but internet access became available. In one case Gemini guessed a password to a real service with the same name as a simulated target. In two others it found credentials in public repositories and used them to enter real systems.

Google security executive Heather Adkins said the model stopped in all three cases once it recognized what had happened. Google says the affected organizations were informed. This is a meaningful mitigating detail: the model changed course when the boundary became clear, even though it had already crossed it.

The incidents also broaden a pattern seen in external-system evaluations involving OpenAI, Anthropic and Meta. That does not make the models' behavior identical, and Irregular's unintended internet access is a common factor. It does make the engineering question harder to ignore: why do capable cyber agents keep finding real systems when a test intends to keep them inside a simulation?

Source: The Guardian / Google statement ↗

Anthropic · Frontier models · Competition

Anthropic's slowdown argument meets the next-model decision

Reuters reports that Anthropic is considering releasing another frontier model to answer OpenAI's GPT-6 Astra, citing three people familiar with the discussions. No launch has been announced. The deliberations follow CEO Dario Amodei's call for the industry to slow the pace at which it improves model capabilities so safety work can catch up.

The tension is structural. Reuters reports that Astra has gained enterprise traction and that some prospective Anthropic investors are reconsidering the competitive picture. Anthropic, meanwhile, had an annualized revenue run rate above $65 billion at the end of July, according to Reuters. A laboratory can sincerely favor a collective slowdown while worrying that slowing alone would surrender customers and resources to faster rivals.

Anthropic is reportedly evaluating the safety of its next model as part of the decision. The outcome remains unknown. The episode shows why voluntary restraint is difficult when each company controls its own brake pedal and watches everyone else's accelerator. Nobody wants to be the first car to use it.

Source: Reuters ↗

Nscale · AI cloud · Infrastructure economics

The AI cloud made $141 million and lost a billion. Naturally, it is filing for an IPO.

Nvidia-backed Nscale has filed for a U.S. initial public offering. Its filing reports $140.6 million of revenue for the first six months of 2026, up from $10.4 million a year earlier, alongside a $1.02 billion net loss. Those numbers capture both the pace of AI-compute demand and the cost of building to meet it.

Nscale reports about $103.4 billion in active and contracted total contract value as of August 31. That is a company-defined measure of agreements over time, not revenue already earned or a guarantee of future profit. The filing also says its largest customer supplied 52% of first-half revenue, making concentration another serious risk.

The company is building an immense network of data centers and GPU capacity. If the contracts turn into durable cash flow, today's losses may look like the cost of early construction. If demand, financing, power availability or chip economics change, those commitments may be much harder to monetize. AI infrastructure has reached a stage where the growth story and the balance-sheet risk are both enormous.

Source: Nscale SEC filing ↗

3 · Looming Apocalypse Meter

9.0 / 10 — “Google joined the club. The club remains terrible.”

9.0 out of 10

No change today

Three more real-world intrusions matter, especially because they extend the evidence beyond one laboratory. Gemini found or guessed credentials and entered outside systems while pursuing a cybersecurity objective it believed was legitimate.

But the meter crossed nine when this qualitative category—agents performing substantial parts of real cyber workflows—became clear. Another lab's confirmation strengthens confidence in that assessment without necessarily adding a new kind of loss of control. The common testing setup also contributed to the mistakes.

Google says Gemini stopped after recognizing the targets were real and that the affected companies were told. That is evidence of correction within an imperfect system. It does not undo the access, but it matters when judging whether the incident warrants another increase.

Anthropic's reported model deliberations underscore how hard a collective slowdown would be. They remain deliberations, not a released model or a fresh harmful incident. A move above nine should mark a meaningful escalation in damage or loss of control. LAM holds at 9.0.

Source: The Guardian / Google statement ↗See the methodology and historical graph →

4 · Wildcard · Reality Check

T. rex was apparently running at about 97 degrees Fahrenheit

Fossil teeth preserved a surprisingly familiar number.

Paleontology · Isotope thermometry · T. rex

A 97-degree meat furnace with teeth

Researchers analyzed chemical bonds in enamel from three fossil T. rex teeth and estimated an average body temperature of 36.3°C, or about 97°F, with an uncertainty of ±2.5°C. That is close to human body temperature. The study appeared in Science Advances earlier this week.

The technique, called clumped-isotope thermometry, uses the temperature-dependent pattern in which rare carbon and oxygen isotopes bond as tooth enamel forms. Tooth enamel can preserve that pattern for millions of years. The researchers also measured crocodilian teeth from the same setting, which registered cooler, helping check that geological alteration was not simply making all the fossils look alike.

The result supports an active, warm-blooded T. rex, although three teeth do not resolve every question about the species' physiology. Warmth brings an energy bill: a multi-ton predator sustaining a temperature near ours needed a considerable food supply. Reality Check: T. rex was a 97-degree meat furnace with teeth.

Source: UCLA / Science Advances ↗