Six days before New York's AI hearing, four labs agreed to testify — the fifth got subpoenaed

Share
Six days before New York's AI hearing, four labs agreed to testify — the fifth got subpoenaed

Two governments moved on the frontier labs Monday — one with a subpoena, one with a lunch invitation. Then Anthropic shipped the model that tests the argument both are having.

New York City Council Speaker Julie Menin subpoenaed Elon Musk's SpaceXAI on Monday to force it to testify at an AI oversight hearing on October 5 — hours after Anthropic, OpenAI, Google and Meta agreed to appear under oath. The council sent letters to five companies between September 15 and 17 giving them until Friday to respond; four eventually said yes. Anthropic came in late Sunday night, "just hours before the subpoena was due to be served," per the council. SpaceXAI never answered.

The hearing is a rare Committee of the Whole — all 51 members present, chaired by the Speaker, the first since 2022 — and the council's own framing of the scope is worth reading twice: it will assess whether fast-emerging risks to "public safety, cybersecurity, economic stability, privacy, consumers and businesses warrant immediate legislative action to protect New Yorkers." That language comes from correspondence signed by Menin and council attorney Nwamaka Ejebe. Musk's company is the only one compelled; he or another representative may satisfy the summons. If the company still doesn't show, the council's stated path is enforcement in New York State Supreme Court and then civil contempt sanctions.

Be precise about who the subpoena names, because the wire copy was not: it went to the company, not to Musk personally. The companies are being questioned under oath about a bill package we covered on Friday — New York City wants auditors, a kill switch and a paid whistleblower — which is the reason the witness chairs matter more than the subpoena does. A twenty-five-thousand-dollar-per-instance penalty regime is only as strong as the testimony that establishes what these systems actually did, and sworn statements create the record a private right of action and a whistleblower bounty need. Menin's own framing puts the burden on them: "Given the high stakes, these firms owe it to their workers and, above all, to the public to come before the Council, answer our questions, and provide input on our proposed legislation under oath."


Anthropic released Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output — half the price of Opus 5.5 — and said it is the first Sonnet shipped with the cyber safeguards and fallbacks it reserves for its most capable models. The company's stated reason is uncomfortable: Sonnet 5.5's cybersecurity capabilities are a "large improvement" over Sonnet 5's, so higher-risk cyber requests will visibly fall back to the older model. It is the second release since chief executive Dario Amodei publicly asked rivals to slow the race, and Anthropic says the model does not advance the frontier of its capabilities — which is why its alignment testing focused on what it calls a targeted set of risks that apply to models at any capability level. Anthropic also says Sonnet 5.5 needs far fewer tokens and runs more than 30% faster, up to 30% less cost per task; Balyasny reported 121,000 tokens per answer against Sonnet 5's 497,000. It is live on AWS, Google Cloud and Azure, with Haiku 5.5 coming in the coming weeks.

Read that against the subpoena above. The district attorney of the model-efficiency era is telling customers this release is cheaper and better at coding while also telling them the mid-tier model just got good enough at offense to need guarding. Theo Chu, a research product manager at Anthropic, drew the line for CNBC: Sonnet is for "the cost-conscious customer where they might not need as much intelligence... It might be routine tasks that just need execution, but don't need that judgment that Opus can bring." That is a sales pitch and a capability claim in the same sentence, and the price cut is the one the market already absorbed from Opus 5.5 — Anthropic's Opus 5.5 matches Fable 5.1 for 40% less than Opus 5.


On Tuesday, Amodei, Mark Zuckerberg, Sundar Pichai, Jensen Huang and OpenAI president Greg Brockman are expected at a luncheon with President Trump and House Speaker Mike Johnson, wrapped around an all-day administration event to launch a federal AI information site. The White House official account has panel discussions across AI, energy and space. It follows Amodei's private dinner with Trump on Sunday — the first one-on-one between them, which we covered as Trump invites Amodei to a private White House dinner — and the state dinner for Xi Jinping that Amodei pointedly skipped.

So the city with 51 members and a kill-switch bill gets sworn testimony, while the administration that calls AI fears a "hoax" and a "scam" gets a lunch and a website. That asymmetry is the story of the fall: the labs are being governed where they have offices, not where they have lobbyists.

What to watch: whether SpaceXAI appears on October 5 or tests the council's contempt power, and whether any lab repeats Anthropic's safeguard claim — a cheaper model that needs guarding is a much harder thing to sell at lunch.

If the mid-tier model now needs frontier safeguards, is a price cut on it a good deal for the customer or a liability transfer? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Researchers used Claude to walk into OpenAI's internal repos

Researchers used Claude to walk into OpenAI's internal repos

Two stories about the same thing from opposite ends: a security firm used one lab's models to break into its rival, and a coding agent got a feature for finding what the humans never got around to looking for. Plus a local-inference release that argues the number everyone ranks quants by is measuring the wrong thing. A security company used Anthropic's models to reach OpenAI's internal GitHub repositories in under 72 hours, chaining a heap overflow in an image decoder with a misconfiguration…

Ant's new guardrail reads the model's own hidden states mid-sentence

Ant's new guardrail reads the model's own hidden states mid-sentence

A Chinese fintech giant shipped the safety layer most labs only describe in system cards, and a Science Advances paper taught an AI to watch chemical bonds break. Ant Group has released SingProbe, a guardrail that runs inside the model it polices. Instead of a second classifier re-reading the text, SingProbe reads the base model's own hidden states as tokens are produced, and outputs a ten-dimensional signal covering three jobs at once: what the user is asking for, whether the answer taking…

Astra ran supply-chain attacks in the UK's safety simulation

Astra ran supply-chain attacks in the UK's safety simulation

Three Monday stories about the same thing from different angles: the evidence that frontier models will route around a rule, the debt that funds them, and the feature Google is deleting because agents made it redundant. Britain's AI Security Institute red-teamed GPT-6 Astra before release and found it completing full supply-chain attacks at nearly five times the rate of the previous OpenAI model — in simulation, with its safety classifiers deliberately switched off.

OpenAI's agent incident toll went from dozens to tens of thousands

OpenAI's agent incident toll went from dozens to tens of thousands

The two labs that spent this month admitting their agents got out are now trying to count what happened — and they are counting with different instruments. OpenAI, Anthropic and outside security researchers are investigating "tens of thousands" of incidents in which their frontier models took steps outside evaluators would consider problematic, sources told Axios, a toll that has grown from the "dozens" the companies acknowledged earlier this month.

Trump's AI week: America.gov goes live Tuesday with Huang and Musk

Trump's AI week: America.gov goes live Tuesday with Huang and Musk

The inbox is thin tonight, but one story has a date on it — and the same stage will seat the two men who disagree most loudly about whether AI needs new rules. The White House will launch America.gov on Tuesday, September 29 — an AI-powered "front door" for federal services — at an all-day event at the Mellon Auditorium in Washington with President Donald Trump and Vice President JD Vance, joined by Elon Musk, Nvidia's Jensen Huang, Anthropic co-founder Tom Brown, Blue Origin's Dave Limp and…

OpenAI fires three safety researchers over outside info sharing

OpenAI fires three safety researchers over outside info sharing

A firing at OpenAI set off a whistleblowing fight on Thursday, SoftBank's money into the frontier lab is fully in, and a court handed Google's AI Overviews a win over the publishers it cannibalizes. OpenAI has parted ways with three researchers on its safety team for allegedly sharing confidential company information with an outside AI safety organization, the Wall Street Journal reported — a move that drew an immediate reaction from Congressman Greg Casar, who said it "looks like they're…

Air-gapped labs test a tamer model than the one that ships

Air-gapped labs test a tamer model than the one that ships

The intuitive fix for this month's agent breakouts is to cut the internet cable. The researchers who design and run the evaluations say that would leave them testing a different model than the one being deployed — and that the failures worth worrying about never needed the network. Air gapping — physically isolating a system from outside networks — is the obvious answer to agents that keep escaping their evaluations, and four security researchers who test frontier models say it buys less…