Today's Headlines

  • METR estimates Anthropic's progress has likely sped up to about twice its earlier pace, with most key evidence redacted
  • OpenAI introduces BioBasicsBench for reading raw lab data and reports 66.7% for GPT-6 Astra
  • Censys finds exposed consoles of the AI penetration testing tool ARTEX on 1,465 hosts
  • Meta pauses a commercial terms update tied to AI training

Today's stories put numbers on where AI stands. METR used Anthropic's internal data to estimate how fast its AI development is moving, and OpenAI tested its own models on reading raw experimental data.

Each number comes with conditions on how to read it. Most of METR's key evidence is redacted, OpenAI says high scores do not yet mean reliable judgment in the lab, and Censys says some of the hosts it found are likely run by authorized testers.

Today's Top Three

METR estimates Anthropic's progress has likely sped up to about twice its earlier pace, with most key evidence redacted

METR, a nonprofit that evaluates frontier AI models, published a report on November 5 saying Anthropic is likely improving its models on well-specified tasks at around twice the rate of March 2024 to October 2025.

The assessment ran from August 12 to October 29. METR says it would not be surprised if the true figure were closer to 1.3x or 3x.

AECI, Anthropic's version of the Epoch Capabilities Index, has grown at an estimated 1.9x the earlier rate since October 2025 (95% confidence interval: 1.5–2.8x). METR describes high uncertainty and internal disagreement on its team.

METR says a majority of the acceleration is likely due to Anthropic's use of AI and guesses the current share at perhaps 80%. It says it is confident that faster growth in training compute did not cause the acceleration. Staff self-reports in Slack polls reached a 12x speedup, which METR treats as only very loose evidence, and it says highly significant, fully autonomous AI contributions to the most important parts of model training have yet to arrive.

METR says it found no clear bottlenecks for the next 12 months and sees a significant risk of even faster progress. It adds that its view of those bottlenecks changed several times during the work and that it is not confident in that assessment.

Anthropic redacted a majority of the evidence METR found most informative. Anthropic described the redactions as "aimed at removing confidential information that could accelerate other AI developers."

METR writes that the public has to rely on its judgment "much more than would be ideal." METR says it took no payment; Anthropic provided laptops and token credits.

OpenAI introduces BioBasicsBench for reading raw lab data and reports 66.7% for GPT-6 Astra

OpenAI introduced BioBasicsBench on November 6, a benchmark of how well AI agents interpret raw experimental data from routine wet-lab work.

The benchmark, developed with the enzyme design company Ligo, has 100 tasks across five areas, including protein methods, microscopy and molecular cloning. Agents work from gel images, plate photographs and instrument exports, and are tested on finding errors and recognizing when the evidence cannot support a conclusion.

In OpenAI's own evaluation at the "xhigh" reasoning setting, GPT-6 Astra earned 66.7% of the available credit, GPT-6.1 Sol 62.0% and GPT-5.6 Sol 43.8%. The run covered seven models and 37 settings, with five attempts on each task, for 18,500 results.

OpenAI's examples include Astra spotting from a rim label that a control plate carried the wrong antibiotic. The company also says "higher benchmark scores do not yet mean models can make reliable judgments in the lab," and shows a case where Astra let an assumption override the controls.

Censys finds exposed consoles of the AI penetration testing tool ARTEX on 1,465 hosts

Censys, which scans the internet for exposed systems, reported on November 5 that it observed exposed consoles of ARTEX, an open-source AI penetration testing tool, on 1,465 hosts in early November.

About 650 of those sit in two blocks of identically configured hosting, which Censys says leaves roughly 815 distinct-looking deployments. Its top countries are the United States at 726, Hong Kong at 387, China at 177 and Japan at 51. Censys says some of the hosts are likely run by authorized testers.

ARTEX was published on GitHub in July. In October, CrowdStrike reported its use in activity against South Korean financial institutions, and Microsoft reported its use by Storm-0940, which it calls a China state-aligned actor, in the compromise of ConnectWise customers.

According to Censys, the developer posted a statement claiming no involvement in the attacks and said the project would go closed-source. The repository is gone, and archived copies still circulate. It also says 155 of the hosts run at least one other attack tool, and that it observed 32 open directories on ARTEX hosts exposing attack kits and similar files.

Other Developments

Product

Research

Policy

Business

Source: AI news inbox (collected November 7, 2026; 13 items, 5 primary and 8 secondary), selected by the editors.