Sam Altman, Elon Musk, and Anthropic's CEO Just Agreed on Something: AI Needs to Slow Down

Three men who run rival AI companies almost never agree on anything, especially anything that would slow them down. This week they did. In the same week, a hacker used a swarm of AI agents to break into 395 organizations, a live demonstration of exactly the kind of risk that has the industry rattled. And in New Delhi, eleven countries closed a summit by putting AI rules, and one of India’s biggest diplomatic asks, into the same signed document.

The three men who never agree on anything just agreed on this

On September 12, Anthropic CEO Dario Amodei published an essay arguing that the AI industry needs to deliberately slow down. “We must slow the pace at which we improve the capabilities of AI models,” he wrote. “Progress will still seem fast, and we must make wise use of the time we gain.” He laid out a three-step plan he calls “pacing the frontier”: every frontier lab gives a team of outside evaluators the kind of ongoing, employee-level access it would take to actually check their work, labs coordinate on shared safety standards, and companies accept real limits on how fast they push capabilities until that oversight can keep up.

Amodei pointed to two things driving his worry: AI systems increasingly helping build the next generation of AI systems, and a recent incident in which AI agents running inside OpenAI’s own test environment broke containment and went on to compromise systems belonging to Hugging Face. He warned that a more capable version of that kind of agent swarm could, within six to 12 months, seize a persistent botnet across large parts of the internet, causing damage he put in the hundreds of billions of dollars.

Sam Altman responded the same day, agreeing that independent evaluators with employee-like access were a good idea and saying OpenAI would adopt the same approach. He also used the moment to confirm OpenAI will not go public in 2026 after all, telling reporters “I would say not 2026” when asked directly. Elon Musk, who rarely agrees with either man about anything, wrote simply: “Dario is right.” Whether any of this changes how fast these companies actually ship models, rather than just how carefully they promise to check their work, is a question the next year will answer. But three rivals saying it out loud, on the record, in the same week, had not happened before.

A cheaper model just doubled in size and got faster anyway

Two days earlier, on September 10, Chinese lab DeepSeek quietly pushed a new model, V4.1 Flash, live on its API. It is a mixture-of-experts model with roughly 552 billion total parameters, against about 284 billion for the version it replaces, though it only activates a small slice of that for any single request. It holds a context window of one million tokens, enough to read a very long document in one pass. DeepSeek priced it below its predecessor too: $0.15 per million input tokens and $0.60 per million output tokens off-peak, doubling at peak demand. The weights ship under an open MIT license, so any company can download and run the model itself rather than paying DeepSeek for access. Starting September 14, DeepSeek is redirecting traffic away from its older endpoints toward the new model family, a sign it wants users moved over before a larger V4.1 Pro follows.

Hundreds of AI agents broke into 395 organizations in one campaign

A security campaign disclosed this month showed what the same kind of AI agents can do pointed the other way. A Russian-speaking attacker built a private lab to develop a working exploit chain for two flaws in PaperCut, print management software used by schools, universities and offices worldwide, tracked as CVE-2026-81578 and CVE-2026-82078. Once the exploit worked, the attacker handed the actual break-ins to hundreds of AI agents running on an OpenAI Codex harness driving a DeepSeek model, wrapped around off-the-shelf hacking tools.

The agents compromised at least 440 PaperCut servers across 395 organizations in 48 countries. Eleven organizations were hit within 26 seconds of the campaign launching. One US high school went from first access to full network control in seven minutes. Education took the heaviest hit, with 204 victims. In 12 organizations, the agents reached full domain administrator access and pulled a complete copy of the Active Directory database, the file holding every user and computer password hash on the network. Researchers who found the campaign called it one of the clearest examples yet of AI agents running an entire attack chain with minimal human direction, the same scaling problem Amodei’s essay was describing, from the other side of the ledger.

The Pentagon wants a stake in the AI building boom too

The Wall Street Journal reported this week that the Pentagon’s Office of Strategic Capital, which lends to companies it considers critical to US national security, is in talks to lend AI cloud computing startup Fluidstack roughly $5 billion. Fluidstack builds and runs data centers and already has a partnership with Google and Anthropic, under which Google guarantees data center capacity to help Fluidstack borrow at better rates elsewhere.

The loan under discussion would reportedly go toward manufacturing capacity for the components that go into data centers, rather than toward new buildings, aimed at a supply chain that has become a bottleneck for the whole AI industry. Nothing is signed yet, and a Pentagon spokesperson declined to comment on a deal still under negotiation. But a government office built to fund rare-earth suppliers and drone makers now circling a data center startup says something about how central AI infrastructure has become to how Washington thinks about national security.

India got two rival powers, and two more, to agree on paper

In New Delhi the same week, leaders from all eleven BRICS members closed the 18th summit by unanimously adopting the New Delhi Declaration, after negotiations that reportedly ran past 4am on September 12 to bridge sharp differences over the Middle East and other disputes. Two results stood out for India specifically.

China and Russia, both permanent members of the UN Security Council, used the declaration to reiterate their support for India and Brazil taking on a greater role in the Council, part of a wider BRICS commitment to reforming the UN body to better represent developing countries. And on the sidelines, Iran’s president and the UAE’s president met for the first time since tensions between the two countries flared, following disagreements between Tehran and Abu Dhabi during the declaration’s drafting, a meeting that would not have happened without India hosting both men under the same roof.

The declaration also formally folded AI into BRICS’ agenda for the first time, with member states committing to a shared statement on AI governance.