OpenAI's 10,000-agent proof raises a customer trust question

OpenAI says an internal model used 10,000 agents over 88 hours to produce a Navier–Stokes singularity proof, after rumors that outside researchers were close. The company says it didn't access their specific user data, but added: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models."

Jacob Coxon resigns over the race to self-improving systems

Jacob Coxon via Techmeme

Coxon resigned Tuesday after three years of pretraining research at OpenAI and Anthropic. He says neither company has met its responsibility.

“The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt. If anything, many executives and senior researchers will couch their phrasing in the press to sound sensible - but I hear the same people express fear privately. No other human activity poses this level of danger.”

Evan Hubinger puts the risk above 10%

CNBC

Anthropic's alignment-science lead supported Coxon's warning and said the lab still lacks a plan for aligning superintelligence.

“Jacob is correct here — we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.”

Samuel Marks says senior employees are more concerned

Axios

Anthropic's scalable-oversight lead joined the public warning while remaining at the company.

“AI developers believe their technology could cause human extinction (or similarly bad outcomes). This could happen in the next few years. In general, the more senior the employee, the more concerned they are.”

US agencies call model distillation an industrial campaign

NSA, CISA, and FBI via Techmeme

A joint advisory says Chinese AI companies, including DeepSeek, are conducting industrial-scale distillation against US frontier models. The accusation turns rate limits and account controls into a federal security concern; it doesn't establish that DeepSeek's new v4.1 flash model was distilled.

Anthropic withholds Mythos 5.1 from the UK AISI

Financial Times via Techmeme

Anthropic declined to provide Mythos 5.1 for pre-release testing, according to the Financial Times. Voluntary evaluation regimes depend on continued lab access, and this refusal exposes that dependency.

A 20-line Lean proof bypasses 13 million lines of formal work

Trail of Bits

Trail of Bits says it proved Fermat's Last Theorem in 20 lines by exploiting a Lean bug, bypassing a 13-million-line formalization. Machine-checked proofs still inherit the security properties of the checker.

Geordie ties AI spending to agents and workflows

SiliconANGLE

Geordie's Cost Intelligence connects enterprise AI bills to the agents and workflows that incurred them. Model and token totals alone don't identify an owner or explain which automated process spent the money.

Euno raises $23 million for agent context

SiliconANGLE

N47 led Euno's $23 million Series A, with seed investor 10D returning. The company is selling a shared context layer for autonomous agents.

Why agents need employee IDs

Forbes Technology Council

Linus Hakansson argues that every production agent needs an organizational identity, not only an API credential. The proposal gives companies a way to assign ownership and answer basic questions about each agent.

Suno replaces its models with v6 trained on licensed music

TechCrunch

Suno says v6 doesn't use the music that trained its prior models and was built through licensing partnerships with Warner Music Group and BMG. The new arrangement pays labels and publishers when the models are used, and users can't prompt for specific artists.

Google commits €13 billion to Finland

CNBC

Google plans its largest European investment to date, spending more than €13 billion over two years. It will build three new Finnish data centers, expand an existing site, and invest about $15.1 billion in dollar terms.

US states pull back data-center tax breaks

Wall Street Journal via Techmeme

More than ten states have rolled back tax breaks for major data-center operators, and others are considering similar bills. In some states, those exemptions exceed $1 billion a year. Losing them changes the economics of where companies build new compute.