Anthropic's models are speeding up its own AI research, METR assessment finds

An independent assessment published on 5 November concludes that Anthropic is likely improving its models at about twice its recent historical rate. METR, a nonprofit that evaluates frontier AI systems, spent eleven weeks inside the company between 12 August and 29 October. Its staff reviewed internal training metrics, benchmark scores, forecasts and interviews with employees.

The report puts the trend in METR’s benchmark at 1.9 times the growth rate from March 2024 to October 2025. Its 95 percent confidence interval runs from 1.5 to 2.8 times. The researchers say the acceleration was not caused by faster growth in training compute, which has long driven AI progress. They also rely on internal Slack polls in which employees reported being about 12 times as productive with Anthropic’s most capable model. That is roughly double what staff reported a few months earlier.

Anthropic was allowed to redact the public version. It removed what METR described as “a majority of the evidence we found most informative.” Cofounder Sam McCandlish said publishing the drivers would hand competitors a roadmap. METR also reports that three models finished training between July and October have not been released. It warns that fully autonomous contributions may soon play a larger role.

Featured source: The Verge , METR .

Other Stories

Chainalysis says attackers took credentials and tokens for every customer

Blockchain analysis firm Chainalysis said an October intrusion exposed names, usernames, password hashes and live tokens for all of its users. Part of the campaign involved an on-path attack that exposed a subset of customer passwords. The company said the actor used AI agent-assisted offensive techniques. Chainalysis took down its login page and know-your-transaction service, then began restoring access with IP allow-listing and mandatory credential resets. It has not said where the traffic was intercepted or whether it was encrypted.

Filed from: Ars Technica .

Nvidia tells foreign cloud customers to earn most revenue outside China

Nvidia now requires data center and cloud customers to draw at least 51 percent of their revenue from outside China, according to the Financial Times. The condition was introduced during the summer and applies to new agreements worldwide, including operators that serve Chinese technology companies. TechRadar, citing the FT, reports that sales representatives and regional managers communicated the rule verbally. Nvidia has not confirmed the requirement publicly.

Filed from: TechRadar .

House chairs call Jensen Huang to a closed briefing on espionage concerns

The Republican chairs of the House committees on China, Foreign Affairs and Intelligence asked Nvidia chief executive Jensen Huang to appear before them. The joint closed-door briefing is set for 18 November. They cited reporting that US intelligence officials believe Huang was targeted by a Chinese spying operation. The reported focus is an executive at Megaspeed, a Singapore cloud company that has bought billions of dollars of Nvidia chips. Nvidia rejected the claims and would not say whether Huang will attend.

Filed from: Ars Technica .

OpenAI raises $11 billion at a valuation of $1.25 trillion

OpenAI has raised another $11 billion at a valuation of $1.25 trillion, Bloomberg reported. A March round valued the company at $852 billion. OpenAI is still seeking investors and had hoped to raise as much as $30 billion at $1.4 trillion before technology stocks fell. The round closed as midterm results showed public opposition to data center projects and as rival Anthropic prepares to go public.

Filed from: The Verge .

Censys finds more than 800 Artex hacking deployments open to the internet

Censys reported more than 800 distinct deployments of Artex reachable from the internet without a password. Artex is an open-source autonomous penetration-testing tool built from AI agents. It has been linked to attacks on South Korean financial companies and to the ConnectWise intrusion Microsoft described in October. The project has since said it will move to closed source, though copies of its code remain available. Censys researcher Silas Cutler said not every exposed instance is run maliciously, though the software does not need internet access to work.

Filed from: TechCrunch .

From the Community

Firecracker escape yields root on the host

Security researchers at PWN.AI describe a chain that breaks out of an AWS Firecracker microVM and runs code as root on the host. The guest locates the Firecracker process through host physical memory. It then recovers its address space despite address randomization and rewrites the saved registers of the thread running the virtual CPU. The proof of concept targeted Firecracker v1.16.1 and ran without its jailer. Deployments using privilege dropping and seccomp would need separate assessment.

Filed from: PWN.AI .

GitHub rebuilds Git storage for thousands of concurrent agents

GitHub described a storage architecture, codenamed Rubicon, that it says can handle thousands of writes per second to a single repository. Internal testing supported 5,000 concurrent agents on one repository and 7,200 pushes per second. Git objects are immutable and content-addressed, so only small reference updates need agreement between writers. That lets storage and request serving scale separately.

Filed from: GitHub Blog .

Anthropic describes how Claude agents maintain their own memory

Anthropic published an account of how Claude agents keep memory between sessions. One process, which it calls dreaming, has the model review its own transcripts while idle. In one evaluation, Claude played Pokémon Red with a folder of notes. An older model wrote 413 files and earned one badge, while a newer one kept nine files and earned four. On an internal corpus of 350,000 files, scaling from two to 128 sessions raised scores on a 25-question evaluation from 13 to 48 percent.

Filed from: Anthropic .