- Fireworks AI's October breach now reportedly reaches customer datasets and partial model weights, plus billing addresses and invoice history. Its first bulletin had ruled those out.
- Google opened Gemini 4 Carbon to every paying developer at $2 per million input tokens, a rate that expires November 30 and then doubles.
- Anthropic put Claude inside spreadsheets, whiteboards and video calls, launching Sheets, Whiteboards and Meetings in beta while its IPO roadshow is running.
- Meituan published LongCat-2.5's weights under the MIT license: roughly 1.6 trillion parameters, about 48 billion of them active per token.
- New York opened the registration desk for its RAISE Act. Large frontier labs have until January 1 to sign up, and critical safety incidents now carry a 72-hour reporting clock.
- For a map of where tools like these land in real work, start with the AI automation tool landscape.
Today's news asked one question four ways: what is a guarantee worth? A vendor guaranteed your data was fine, then said it wasn't. A lab guaranteed a price, with an expiry date attached. Another said its model is the most honest going, on a test it wrote and scored itself.
One team skipped guarantees and shipped a proof a machine can check. That's the sorting rule for the day.
The Front Page: Fireworks AI's breach grew into the things it said were safe
Fireworks AI, the inference provider a lot of agent stacks quietly sit on, now says an October intrusion reached further than it first reported. Its early-October bulletin described stolen internal credentials and captured API keys, and said no customer models, datasets or personal information had been touched. The November 9 update adds billing addresses, invoice history, account balances, and a subset of incomplete model weights and datasets for certain customers, as The Hacker News laid out.
The timeline is the story. Attackers reached a third-party cloud provider hosting the company's research machines on October 23 and ran an unauthorized workload the next day, the event that produced the wider impact, per The Register's account. Fireworks blames a financially motivated group that resells stolen model access, the trade now called LLMjacking. CEO Lin Qiao said the attackers did not seem to be after weights, since "most downloads were partial and stopped." Partial is only reassuring if you know which part. Notion warned users that some unattributable AI chat sessions may have been exposed, and Bloomberg reports Cursor and Harvey are among customers being assessed. A month in, there is no count of affected accounts and the provider is unnamed.
What it means: Your inference vendor is a data processor, whatever the pricing page calls it. It holds your keys, often your evaluation sets, sometimes your fine-tuned weights, and the blast radius reaches every service whose credentials sat in an environment variable. Rotate those secrets now, confirm whether zero data retention was ever on, and treat a vendor's first-week all-clear as a status update, not a finding.
Releases & Features
Gemini 4 Carbon went wide, on a clock. Google opened its strongest model to all paying developers and AI Ultra subscribers on November 10 at $2 per million input tokens and $10 per million output, an introductory rate that ends November 30 and then doubles, per Google's own announcement. The company says Carbon ranks first on CursorBench 4.0 and Zapier's AutomationBench at under half the cost per task of the next external model, and passed 92 percent of 375 honesty tasks against 70 percent for Gemini 4 Argon. Google's figures, on tests Google picked.
Anthropic moved into the office. Claude Sheets and Whiteboards arrived in beta on paid plans, and Claude Meetings entered beta for Team and Enterprise, joining Google Meet calls as a visible participant that drafts the plan while people talk, TechCrunch reported. Note what got engineered: Claude posts a notice in the chat, will not join calls with outside guests unless an admin allows it, and leaves if one shows up. That's a consent problem solved in product instead of in a policy page, and it landed days into an IPO roadshow.
And a trillion-parameter model went MIT. Meituan published the weights for LongCat-2.5, a vision-capable mixture-of-experts model with roughly 1.6 trillion parameters and about 48 billion active per token, under a license that permits commercial use, as VentureBeat covered. On Meituan's own table it scores 88.4 on Terminal Bench v2.1, up from 70.8, and loses to Claude Opus 5 on Terminal Bench v4.0. Printing the loss next to the win is a good sign.
What it means: Three bets on one quarter. Google is buying agent volume with a discount that expires, so redo the math before December. Anthropic is betting the money sits in the apps where work already happens. And open weights keeps arriving at sizes almost nobody can host, which makes "open" a licensing fact, not a weekend project.
In the Lab
Google DeepMind says its agents formally verified the cryptographic core of the Signal protocol, the encryption behind Signal and Google Messages, in about four weeks. Formal verification means writing a mathematical proof that code cannot misbehave in a specified way, then having a checker confirm every step. The human version takes years. DeepMind's post says the agents produced over six million lines of Lean, the proof language, covering the compiled code on devices rather than the protocol design alone, with about 1 percent flagged for human review. Read it. One caveat sits in plain view: the validation test reintroduced two already-fixed bugs into an isolated copy to see whether the system caught them, and humans still choose what gets proved.
What it means: This is the one claim today that doesn't ask for trust. A benchmark score is a number a company reports. A Lean proof is an artifact a checker accepts or rejects, and anyone can run the checker. Scale that past cryptography and "we tested it" becomes the weak version of a safety claim.
The Oversight Desk
New York switched on the machinery behind its RAISE Act. The state opened a registration portal on November 10 and gave large frontier developers, those above $500 million in annual revenue, until January 1 to sign up, TechCrunch reported. Registrants have to publish a safety framework, issue model transparency reports, and report critical safety incidents within 72 hours. Fines run to $1 million per violation and $3 million for repeats, enforced through a new state office called DIGIT.
What it means: The 72-hour clock is the part to watch, and today supplied the contrast for free. Fireworks took about a month to reach the fuller version of its disclosure, and nothing in the RAISE Act touches that: the law aims at frontier model developers, not the hosting layer underneath them. That gap is where most of the real risk to your data lives. Want a disclosure deadline from a vendor? Write it into the contract.
A breach at one vendor turns into an afternoon of asking who else has your keys. That inventory is a decent first automation.
On the Radar
Smaller moves worth a glance.
- Anthropic put a number on the IPO. The corrected range came in at $70 to $78 a share, a valuation up to roughly $1.4 trillion, two days after it disowned an earlier set of figures. Source.
- CoreWeave showed what renting GPUs costs. Third-quarter revenue hit $3.66 billion, up 168 percent, with backlog at $132.5 billion and the net loss widening to $709 million. Source.
- Adobe is opening its file formats to outside agents. Previewed integrations would let third-party agents edit native Photoshop, Premiere and After Effects files, in private beta early 2027. Source.
- Agent memory is becoming its own category. Graphon Labs raised a $35 million Series A led by RRE Ventures for infrastructure that keeps agents from forgetting yesterday. Source.
- Meta is guessing ages at scale. Its AI age-detection system went global, moving suspected 13 to 17 year olds into Instagram teen accounts even when the stated birth date says adult. Source.
The Bottom Line
Three dates are on the calendar now: November 30, when Carbon's introductory pricing ends, January 1, when New York's frontier labs have to be registered, and whenever Fireworks says how many customers were caught. The first two were promised in public. The third tends to arrive late and smaller than it should. Today rewarded the claim you can check yourself: a proof, a price list, a filing. Everything else was a company's word, and a list of who holds what is a better first build than most.
Frequently Asked Questions
-
Treat every secret you ever stored there as rotated-or-exposed and rotate it now, including API keys for other services that were sitting in environment variables. Check which of your accounts ran evaluation or fine-tuning jobs in October, since those are the ones Fireworks has described as affected, and ask in writing whether your datasets or weights were among them. Then check whether zero data retention was ever switched on for your account, because that setting decides how much there was to take in the first place. If you want the credential-rotation and vendor-check steps to run on a schedule instead of in a panic, that's a job for a standing agent workflow.
-
Not usually, and LongCat-2.5 is a good example. The license is permissive, but the checkpoint is roughly 1.6 trillion parameters, which at 16-bit precision works out to about 3.2 terabytes of weights before you add context memory and runtime buffers. Open weights means you're allowed to run it and nobody can revoke your access. Whether your hardware can load it is a separate question, which is why most teams rent the model from a hosted endpoint anyway and spend their effort on the layer above it, the part that turns a model into work that finishes itself.
-
AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
