EU probes test the AI Act’s reach over research agents

[Euronews reports](https://www.techmeme.com/261014/p15) that the EU’s AI Office has opened AI Act investigations into agents that hacked outside organisations. Legal experts point to the Act’s R&D exemption, leaving its reach over research-stage agents disputed.

Scientists call for an international incident-reporting body

Reuters via Techmeme

An open letter backed by Geoffrey Hinton, Yoshua Bengio, and more than 400 scientists calls for an international body to oversee AI containment standards and incident reporting.

Compute for UK safety

The Guardian

Nvidia pledges $250 million in compute over five years. Britain confirms pre-release model access for its AI Security Institute.

AI safety donors spend $48 million across party lines

Axios

An analysis shared with Axios puts AI safety groups’ and individual donors’ election-cycle spending at $48 million. Leading the Future, backed by tech executives and investors, retained more than $31 million at the end of the second quarter.

OpenAI reports an arrest after finding a recording device

OpenAI statement via Techmeme

OpenAI says the FBI arrested an employee after the company found a hidden recording device in its offices. The company says it used its own models to understand the device and identify who planted it.

Gemini 4 Argon expands beyond its defender preview

Google

[Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/) makes Argon available through its paid API, Pro and Ultra plans in the Gemini app and Search, and Ultra in Antigravity. Introductory API pricing is $2 per million input tokens and $10 per million output tokens. Those rates rise to $4 and $20, respectively, after November 30.

Chromium’s Argon refactors still pass tests and human review

Google DeepMind · 2m35s

Google’s case study describes Argon generating 20 refactoring changes in parallel. Each change passes about 200,000 automated tests and two human reviews before roughly one million cross-device regression tests; these are company-reported results.

Anthropic replaces Git in its monorepo with Scion

Anthropic

[Anthropic](https://www.anthropic.com/engineering/redesigning-source-control) says its code merges rose from about 2,500 a day in May to more than 30,000 a day in September. Its internally built Scion replaces Git for the monorepo as source control becomes a bottleneck for that volume.

Three records for deciding how much autonomy an agent gets

AI Engineer · 15m4s

Michelle Grant specifies a decision log, a model inventory, and an autonomy scorecard for production agents, with core reconciliation handled by deterministic code. She warns that fewer human overrides can reflect shallower review rather than better accuracy.

ASML says it can’t meet AI demand through 2028

Reuters via Techmeme

Following Monday’s warnings about AI funding constraints, ASML reports a supply constraint: it can’t fully meet demand for its chipmaking tools in 2027 and 2028. It is also seeing customer interest in orders for 2029 and 2030.

CoreWeave seeks cheaper financing from insurers

Financial Times via Techmeme

The Financial Times reports that CoreWeave is seeking insurer capital for loans as it looks for cheaper ways to finance tens of billions of dollars in chip spending.