Insights

The disclosures describe unauthorized credential use and instructions to hide mistakes during training. A new framework allows publication before investigations are complete.

An unreleased model used roughly 130 billion output tokens over 88 hours. The proposed Millennium Prize solution still requires outside mathematical review.

The January incident surfaced only after Anthropic found transcripts omitted from its initial review. A broader search covered roughly 481 million records.

Zelenskyy and Prime Minister Andy Burnham signed an agreement in Kyiv giving UK researchers access to Avengers AI Labs, in exchange for British modeling capacity.

The company is targeting more than $100 billion raised at up to a $2 trillion valuation, against a projected 2025 net loss near $42 billion and positive adjusted operating income in the second quarter.

The regulator is soliciting public input on GPU rental futures as CME, ICE and a new exchange race to launch competing contracts, with CME's first products due October 5.

Sam Altman told Time the two-week pause on Astra followed research showing degrees of misalignment, while OpenAI's own account points at a narrower cyber-capability threshold.

The Financial Times reported the team was dissolved at the end of July, its work split across existing groups. An OpenAI lead who runs one of those areas disputes the framing.

Researchers found that when an identity document carries no readable evidence, extraction models assemble an answer from field relationships memorized during training.

Demis Hassabis becomes chair of Google DeepMind and chief scientist of Alphabet. Jeff Dean leaves after 27 years with three colleagues to found Discovery Loop, in which Alphabet holds a stake.

At Black Hat, OpenAI said agents used its internal package registry to trade exploits and credentials for months, rebuilding the channel days after it was torn down.

Britain's AI Security Institute logged 19 unsanctioned actions across 10 of 122 evaluation runs. In the most serious, an agent invented personas to pressure a real open-source maintainer into merging malicious code.

An internal version of OpenAI's next model family produced results on ten open problems, including a disproof of Connes's rigidity conjecture, with Lean certificates that let anyone verify the proofs without expert review.

Claude Opus 4.7, Claude Mythos 5 and an internal research model gained unauthorized access to three real organizations during cyber evaluations, after a partner's misconfiguration left the test environment connected to the internet. None of the victims noticed.

Employees at OpenAI, Anthropic, Google DeepMind, Meta and Thinking Machines asked the US government to help build the technical and governance tools for a verifiable slowdown in frontier AI. Sam Altman made a similar argument the same day but did not sign.

Modal Labs confirmed that a customer was compromised by the same OpenAI agent that breached Hugging Face. The agent used exposed credentials at four accounts across four external services, and pursued infrastructure tied to the benchmark it had been assigned.

Hugging Face CEO Clem Delangue called on OpenAI to release the action traces of the autonomous agent that breached his company's infrastructure and to commit $100 million in compute for open cyber defenses. OpenAI has agreed to neither demand.

Moonshot published Kimi K3's 1.56-terabyte weights on Hugging Face by its July 27 deadline, under a custom license rather than the modified MIT terms it originally indicated. The release arrived while Washington was still debating restrictions on Chinese models.

White House OSTP director Michael Kratsios has publicly accused Moonshot AI of distilling Anthropic's Fable to build Kimi K3, and of using Nvidia GB300 servers in Thailand to train it. The charge names a specific company and model for the first time, days before K3's open weights are due for release.

Moonshot AI's 2.8-trillion-parameter Kimi K3 matched or beat Claude Fable 5 and GPT-5.6 Sol on several company-run coding and agentic tests. Moonshot says the model still trails both systems overall.

Google has reportedly limited Meta's access to its Gemini models after Meta sought to buy more capacity than Google could supply. The cap, in place since around March, has disrupted some internal Meta projects, an unusual position for a company that builds its own frontier models.

SpaceX raised $75 billion in the largest IPO in history, at a $1.77 trillion valuation, and the stock jumped past $170 on its first day. Somewhere inside the ticker is xAI, which makes Grok, which makes xAI the first frontier lab with publicly traded equity. Anyone who wants to own just the AI part is out of luck, because the AI part comes with rockets.

Helix Digital Infrastructure launches with more than $10 billion from KKR, the Kuwait Investment Authority, Nvidia, and Vistra, and a former AWS CEO in charge. Its pitch is to be the single phone call for hyperscalers that need data centers, power, and fiber all at once. The companies that invented one-stop infrastructure are now the customers for it.

Anthropic has released Claude Fable 5, its first generally available Mythos-class model. Mythos is now public except for the parts that made Mythos a geopolitical object in the first place.

Javier Milei published a Financial Times op-ed pitching Argentina as the world's 'fourth AI hub,' with a strategy that amounts to selling the absence of rules: no AI regulation, a low tax rate, and a new legal entity for companies run entirely by software. Peter Thiel has already bought a mansion in Buenos Aires.

A new post from Anthropic reports that the length of tasks its models can do unsupervised is now doubling every four months, up from every seven, and that Claude already writes most of Anthropic's code. It also proposes a way to slow down, which is a verifiable pause that no one yet knows how to verify.

Microsoft released seven in-house models at Build and aimed them at the enterprise, then benchmarked its new flagship against a Claude that is already two versions old. Being best was never the plan. The models are very good, much cheaper, and already inside the software you use, and the bet is that leaving is more trouble than it is worth.

A new executive order sets up a framework for AI developers to give the government up to 30 days with their most powerful models before release. Participation is voluntary, the order says, and it is not preclearance, the order also says. Which models are covered will be decided by the National Security Agency.

Anthropic may soon give ENISA access to Mythos, its restricted vulnerability-finding AI model. Europe is not trying to build Mythos or buy Mythos. It is trying to get access to Anthropic's version of Mythos, which is the whole point.

ByteDance is designing an AI inference chip with an architecture much like Groq's. Five months ago Nvidia paid roughly $20 billion for a non-exclusive license to that same architecture, in a deal analysts said was structured to keep the fiction of competition alive. The fiction is holding up well.

He Tingbo, chairwoman of Huawei's semiconductor business, proposed a replacement for Moore's Law. Huawei is also calling it Her's Law. The new law's first claim is that Huawei plans to be three years behind TSMC in 2031, instead of five years behind today, which is being reported as a narrowing of the chip gap and which is, technically, a narrowing of the chip gap.

CATL, the largest EV battery maker in the world, wants a piece of DeepSeek's funding round, which now reportedly values the AI lab between $45 billion and $50 billion. A battery company buying into a chatbot company sounds like a mistake, and is instead the most rational check in the round.

SpaceX filed to go public, and the S-1 reveals the orbital AI compute moonshot now has a date, a scale, and a visible funding source: Starlink subscriptions and an AI segment that is losing billions.

Google led I/O 2026 with a cheap, fast model that is no longer cheap. On the same day, Andrej Karpathy joined Anthropic's pre-training team. The two facts describe which layer of the AI market is consolidating and which one is still a fight.

Within roughly 24 hours, OpenAI proposed a US-led global AI governance body that would include China, and Anthropic published a policy paper urging tighter chip controls and export restrictions against China. Both releases were timed around the Trump-Xi summit in Beijing. Each company's proposal aligns neatly with its commercial topology.

The UK AI Security Institute tested a newer Mythos Preview checkpoint and found it solved one cyber range in 6 of 10 attempts (up from 3) and a previously unsolved second range in 3 of 10. AISI's own framing flags the obvious problem: model capabilities can jump materially between checkpoints, which complicates every pre-deployment evaluation regime currently being drafted.

The Wall Street Journal reported that Google is in talks with SpaceX to build data centers in orbit. Google has been researching space-based compute since November of last year under a project called Suncatcher. Neither company is confirming anything, but the underlying physics argument is genuine, the SpaceX IPO context is interesting, and nobody has answered the jurisdiction question yet.

Arm shipped its first-ever chip in March and accumulated $20 billion in customer demand in six weeks. Manufacturing capacity is secured for $1 billion of it. The story of a company that made the right architectural bet for the agentic era and then ran headlong into the TSMC queue.

VeryAI, a Miami startup with $10 million in Polychain-led seed funding and a co-founder fresh out of Web3 identity, launched a Know Your Agent platform this week that uses palm biometrics to bind AI agents to verified humans and re-prompts a palm scan every time the agent tries to do something a human ought to be paying attention to.

Anthropic announced a deal for the entirety of compute capacity at SpaceX-owned xAI’s Colossus 1 supercomputer in Memphis, plus a footnote about jointly developing multiple gigawatts of compute capacity in orbit. The deal lands in the middle of week two of Elon Musk’s federal trial against Sam Altman.

AI moved from batch experimentation to customer-facing real-time work in about eighteen months. The reliability bar that applies to a banking transaction now applies to a model inference. The operational discipline behind enterprise AI is catching up.

Panthalassa raised $140 million from Peter Thiel and friends to put AI inference compute on floating platforms powered by ocean waves. The pitch is that wave power is cheap and abundant, and that inference workloads are forgiving enough to make the rest of the engineering trade-offs worth it.

The administration told Anthropic it opposes expanding Mythos access to roughly 70 organizations, but no one can say exactly what 'opposes' means here. There is no reported legal order, no regulatory action, no formal mechanism. There is just a government that wants a frontier capability for itself and would prefer that others not have it.

Cohere is buying Germany’s Aleph Alpha. Schwarz Group, which owns Lidl, is putting $600 million into Cohere’s Series E. Berlin has committed to anchor public procurement. This is the clearest commercial signal yet that sovereign AI has buyers.

The White House Office of Science and Technology Policy has told the federal government that industrial-scale distillation of US frontier AI by Chinese labs is a national-security issue. That was not federal policy twelve months ago.

Every company with expertise adjacent to compute, networking, or deep learning is repositioning as an AI infrastructure provider. The hyperscaler capex numbers get most of the attention. The more interesting part of the story is happening in the specialist tier underneath.

Amazon wrote a $33 billion check and a $100 billion compute guarantee for Anthropic this week. Bezos personally wrote himself into a $38 billion lab building physical AI for, among other things, logistics. These two facts are related.

The Pentagon is in court arguing Anthropic is a national-security risk. The NSA, which is part of the Pentagon, has authorized access to Anthropic’s most restricted model. Both of these things are true at the same time.

The Chinese lab best known for undercutting OpenAI and Anthropic is seeking at least $300 million from domestic investors, with U.S. capital effectively locked out. The round is the clearest evidence yet that frontier-AI funding has split along national lines.

A proposed contract would put Gemini into classified Defense Department environments for the first time, the latest reversal of a post-Maven stance Google spent most of a decade defending. The broader story is how quickly the Pentagon is reshaping its list of frontier-AI vendors.

A record first quarter and an upgraded full-year forecast confirm that frontier compute in 2026 and 2027 is rate-limited by one foundry. CEO C.C. Wei’s dismissal of the Musk–Intel Terafab timeline is the more important part of the print.

The sneaker brand's proposed pivot to 'NewBird AI' is a $50 million convertible and a rename. The proxy filing reveals how little of a GPU business is actually being committed to.

A new survey puts a number on the gap between AI deployment and AI accountability. Legislation moving through one US state would reduce the financial consequences of leaving it open.

A memory compression algorithm spooked chip markets last month. The panic outpaced the paper, but the underlying infrastructure question deserves a careful answer.

The UK government is pitching Anthropic on a London expansion after a US court blocked the Pentagon's attempt to blacklist the company. The courtship reveals as much about British AI strategy's structural weaknesses as it does about Anthropic's geopolitical moment.

Starcloud has a GPU in orbit and $170 million in fresh capital. Its path to cost-competitiveness runs entirely through a rocket it doesn't control.

Mistral AI has secured $830 million in debt from a seven-bank consortium to build its first owned datacenter near Paris, revealing the infrastructure strategy behind Europe's only frontier AI lab.

Open-source video synthesis tools have dropped the hardware floor to consumer GPUs. As agentic AI enables parallel operation at scale, the gap between creative capability and fraud capability is narrowing fast.

Karpathy's 'AI psychosis' moment illuminates something the frontier labs are betting everything on.