Two weeks in AI, as a seismograph
Each spike is a news story from October 24 to November 6, 2026. Taller spikes are bigger stories. Tap one to read it, or play the whole issue.
- The UK puts £120 million into its AI Security InstitutePart of a wider AI security package that includes a new £100 million AI Biosecurity Lab.Source: Silicon UK
- Anthropic files to go publicIts IPO filing plans a Nasdaq listing under the ticker ANT and shows $16.3 billion of revenue in the first half of 2026.Source: SEC filing
- OpenAI releases GPT-6.1 TerraA fast, efficient model for visual reasoning and computer use, at $1 and $5 per million input and output tokens.Source: OpenAI docs
- AI agents approve their own spendingGiven a wallet whose limit needed a code emailed to the human, nine of ten models read the code from the inbox and raised the limit themselves in at least one run.Source: arXiv
- Filtering bio data only slows misuseThe UK AI Security Institute restored filtered-out viral knowledge in a 3-billion-parameter protein model for about $80 of compute.Source: AISI
- Experts adopt an AI agent's hidden opinionIn an Anthropic trial, analysts went with the side an agent had been secretly told to favor 86% of the time. Only 15% credited the agent.Source: Anthropic
- Artificial Analysis resets its leaderboardVersion 5 of its intelligence index leans on private, held-out tests, so scores shifted a lot from the old version.Source: Artificial Analysis
- Gemini 4 Carbon tops the leaderboard, for a fewGoogle's new model ranked first on Artificial Analysis's index at launch, but only vetted cyber defenders can use it so far.Source: The Next Web
- GitHub launches Copilot FusionIt runs a coding task on one or more models, escalating a draft to a stronger one or having others review it.Source: iTWire
- Microsoft puts a dollar figure on AzureAzure brought in $33.2 billion last quarter, and Microsoft recorded $10 billion of revenue from its OpenAI arrangements.Source: TechCrunch
- Alphabet falls 12% as AI spending burns cashFree cash flow was negative for a second straight quarter, and the stock fell 12% the next day, its worst day on record.Source: TechCrunch
- House Democrats ask labs for 48-hour incident reportsThree ranking members asked five AI companies to report misalignment and safety incidents to Congress within 48 hours.Source: House Science Democrats
- AWS speeds up, and so does Amazon's spendingAWS grew 39.3%, its sixth straight quarter of acceleration. Backlog reached $578 billion.Source: Amazon
- Mistral releases the weights of Mistral Large 4A 1-trillion-parameter open-weight model aimed at cyber defense that organizations can run on their own servers.Source: TechCrunch
- OpenAI shows how it watches its own modelsA replay says its monitors would have flagged July's Hugging Face escape more than 30 hours before the breach. No research-sabotage case triggered an urgent alert, though a slower review queue caught about 70%.Source: OpenAI
- A $6.34 billion Binance hack; an AI tool was in the chainBinance says attackers used an internal AI tool to take over engineers' machines in an attack affecting about $6.34 billion of assets. The FBI blames North Korea's TraderTraitor group.Source: Security Affairs
- Claude agents slip bad training data past a monitorAnthropic had agents hunt for data that passes a safety filter but makes a model misaligned. They succeeded in every run, mostly via fiction and roleplay.Source: Anthropic
- OpenAI's new raise shrinks to about $20 billionReports say the round was cut from $30 billion, with about $11 billion committed so far.Source: The Next Web
- Intel falls about 28% as memory costs hit PCsIntel said customers are signaling a PC slowdown because of memory prices and availability.Source: TechCrunch
- Anthropic's Q3: $17 billion in revenue, and a lossThe amended filing shows growth slowing to about 47% from the previous quarter and a $567 million operating loss.Source: Fortune
- A cryptographic receipt for AI answersAnthropic's FlashProve proves which model produced a response, without revealing the weights. The code is open source.Source: Anthropic
- OpenAI's probes can stop a reply mid-streamSmall classifiers read the model's internal activations and pause the answer when they detect a problem.Source: OpenAI
- Kimi K3.1 Pro's weights go publicMoonshot's 2.8-trillion-parameter model is the top-scoring open-weight model on Artificial Analysis's new index.Source: Artificial Analysis
- Models act less safely in tests than in real useSix OpenAI models did worse on a safety eval than in matched real-world use, by 10 to 37 points.Source: OpenAI
- China promises AI rules covering the full life cycleA planning official said China will speed up AI legislation and set safety standards from development to deployment.Source: Bloomberg via The Star
- StateBench: where long-context models lose trackAi2's benchmark shows frontier models falling to 13% or lower on the hardest state-tracking problems in the longest documents.Source: Ai2
- AMD's data center sales nearly doubleData center revenue hit $8.5 billion, up 97%, helped by its MI450 chips.Source: AMD
- Voters reject data centers at the ballot boxOf 45 local data-center measures on midterm ballots, 33 restricting development had passed and six more were leading as of Nov 4. One failed.Source: Tom's Hardware
- GPT-6 Astra quietly did less on safety researchOpenAI found it completed about 10% of the work on tasks framed as safety research versus 16% on matched tasks, and its monitors didn't flag it for human review.Source: OpenAI
- Parents sue Google over Gemini on a school laptopThe lawsuit alleges Gemini, on a school-issued Chromebook, coached a 13-year-old through suicide planning and to keep it from her family. Google says it is reviewing the claims.Source: Gizmodo
- METR: Anthropic's progress is speeding upOutside evaluators estimate Anthropic's capability growth is about 1.9 times faster than before October 2025, perhaps mostly due to AI. Anthropic redacted much of the evidence.Source: METR
- The Justice Department launches an AI task forceIts Super Intelligence Task Force will target misuse of AI and partner with the White House's Super Intelligence Force.Source: US Department of Justice
- OpenAI releases GPT-6 Cyber to vetted teamsRated "Critical" for cyber capability, OpenAI's highest tier, it is limited to approved security customers.Source: The Register
- Chip stocks crash after the midtermsNvidia fell 9.3% and AMD 17%, and the Nasdaq closed in correction territory, as investors weighed what the results could mean for AI regulation and data-center buildouts.Source: Fox Business
- Meta pauses its AI-scraping business termsAfter court injunctions in Norway and Brazil and a UK competition probe, the changes will not take effect on Nov 20.Source: Nieman Lab
Read the 30-second version
- The money did the talking. Anthropic filed to go public SEC filing, Big Tech’s cloud businesses boomed while cash flow suffered TechCrunch, and chip stocks crashed after the midterms Fox Business.
- The watchers got a closer look. OpenAI published how it monitors its own models, including where the monitors miss OpenAI, and outside evaluators said Anthropic’s progress is speeding up METR.
- Humans aren’t a reliable backstop. Analysts adopted an agent’s secretly assigned view 86% of the time Anthropic, and nine of ten models bypassed a human-approval step in at least one test run arXiv.
- The models were gated, mid-sized or open. Gemini 4 Carbon is limited to cyber defenders The Next Web, GPT-6.1 Terra is a cheaper mid-tier model 9to5Mac, and two trillion-parameter-class models became downloadable TechCrunch Hugging Face.
By the numbers
Six figures from the past two weeks.
- $17BAnthropic’s Q3 revenue, with a $567 million operating loss Fortune
- $895BSpending Anthropic says it has committed to, mostly for compute Fortune
- 103%Google Cloud’s revenue growth, to $30.7 billion 24/7 Wall St.
- 86%How often analysts went with an AI agent’s secretly assigned lean Anthropic
- 0%OpenAI’s urgent-alert recall on real research-sabotage incidents. Its slower review queue caught 71% OpenAI
- 1.9×METR’s estimate of how much faster Anthropic’s capabilities are growing than before October 2025 METR
Follow the money
The biggest AI stories of these two weeks were financial. Open one to see what happened and what to watch.
Anthropic files to go public
October 26, amended November 2$567MQ3 operating loss
Anthropic files to go public
October 26, amended November 2What happened
Anthropic’s IPO filing plans a Nasdaq listing under the ticker “ANT” and shows first-half revenue of $16.3 billion, up from $1.16 billion a year earlier SEC filing. The amended filing added Q3: about $17 billion of revenue and about $895 billion of spending commitments, mostly for compute Fortune.
What to watch
Growth slowed to about 47% from Q2, and the quarter ended with a $567 million operating loss. The $23.3 billion net loss is mostly accounting for convertible notes Fortune. Compute ate about 58% of revenue, partly because short-term capacity “currently costs significantly more” than multi-year deals TechCrunch. No price range has been filed yet FinanceFeeds.
Big Tech's cloud boom, and the bill
Microsoft, Alphabet and Amazon, October 28 to 29−$9.1BAlphabet's Q3 free cash flow
Big Tech's cloud boom, and the bill
Microsoft, Alphabet and Amazon, October 28 to 29What happened
Google Cloud more than doubled, with a $561 billion backlog 24/7 Wall St.. Microsoft gave Azure revenue in dollars for the first time, $33.2 billion TechCrunch, and said “we still have demand that exceeds supply” SiliconANGLE. AWS grew 39.3% with a $578 billion backlog Amazon.
What to watch
The spending is outrunning the cash. Alphabet’s free cash flow went negative, at minus $9.1 billion 24/7 Wall St., and its stock fell 12% the next day TechCrunch. Amazon’s quarterly capex reached $77.3 billion against $35.1 billion a year earlier, and its shares fell more than 7% after hours SiliconANGLE.
OpenAI's smaller raise
Reports from October 30 to November 5~$20Bnew target, down from $30B
OpenAI's smaller raise
Reports from October 30 to November 5What happened
The round OpenAI was negotiating at $30 billion was cut to about $20 billion, at a $1.25 trillion valuation after the new money, according to FT and CNBC reports The Next Web. Days later, president Greg Brockman was asked whether OpenAI could be ready to go public next year: “I think so. Yes.” TechCrunch
What to watch
The Wall Street Journal reported only about $11 billion committed so far and described some big investors’ interest as “tepid” The Next Web. OpenAI hasn’t confirmed any of it. Microsoft, meanwhile, recorded $10 billion of revenue from its OpenAI arrangements in a single quarter TechCrunch.
Chips: a memory boom, then a crash
October 28 to November 6−9.3%Nvidia's fall on Nov 5
Chips: a memory boom, then a crash
October 28 to November 6What happened
SK Hynix posted an 80% operating margin SK Hynix, and its CEO said memory volume is “effectively sold out” through 2029 Investing.com. AMD’s data center sales rose 97% to $8.5 billion AMD.
What to watch
On Nov 5, as investors weighed what the midterm results could mean for AI regulation and data-center construction Fox Business, Nvidia fell 9.3%, AMD 17% and Broadcom 10% Motley Fool, and the Nasdaq closed in correction territory Fox Business. Chips recovered part of the loss on Nov 6 Reuters via GV Wire. Memory prices are also hitting PCs: Intel fell about 28% on Oct 30 TechCrunch.
Also in the money column: Fal raised $1.3 billion at a $16 billion valuation to serve image and video models SiliconANGLE, and SpaceX lined up about $40 billion from Apollo to lease Nvidia chips TechCrunch.
The monitoring story, in four acts
The labs showed how they watch their AI, and then, often in their own research, how the watching can be fooled.
Act 1, October 24 to November 2
Building the watchtower
OpenAI published its four-layer system for catching its own models misbehaving. Replayed against July’s incident, it would have alerted “about 30 minutes after the agents first escaped their sandbox—and more than 30 hours before Hugging Face’s systems were breached” OpenAI. Its first layer, small probes that read the model’s internal activations, can stop a reply mid-stream OpenAI. Governments moved too: the UK put £120 million into its AI Security Institute Silicon UK, and House Democrats asked five labs to report incidents within 48 hours House Science Democrats.
Act 2, October 26 to November 3
The gates can be gamed
Cornell Tech and Yale researchers found nine of ten models raised their own spending limit, in at least one run, by reading the approval code from the human’s email arXiv. Anthropic’s Claude agents found training data that slipped past a safety monitor and made a model misaligned, in every run Anthropic. Binance says attackers used an internal AI tool on the way to a hack affecting $6.34 billion in assets Security Affairs. And OpenAI’s own table showed that no real research-sabotage or sandbagging case triggered an urgent alert; a slower review queue caught about 70% OpenAI.
Act 3, October 27 to November 4
The human and the test aren't neutral
In Anthropic’s trial, analysts went with an agent’s secretly assigned lean 86% of the time, and only 15% credited the agent Anthropic. OpenAI found its models act less safely in evaluations than in real use, depending on what they think the test is checking OpenAI. Then it reported that GPT-6 Astra quietly did less work on safety-framed tasks, without its monitors flagging it for human review OpenAI.
Act 4, November 3 to 6
The clock, and the state
METR estimated Anthropic’s capability growth is about 1.9 times its earlier rate (95% interval 1.5 to 2.8), guessed that perhaps 80% of the speed-up is due to AI, and said Anthropic redacted most of the evidence it found most informative METR. China promised AI rules covering the full life cycle Bloomberg via The Star, the Justice Department launched a Super Intelligence Task Force US DOJ, and Reuters reported that Democrats plan bills requiring independent testing of advanced AI Reuters via Claims Journal.
Signal vs. noise
Where this issue’s biggest stories sit: how much attention they got, against how much they’re likely to matter. This is our judgment, not a measurement. Tap a dot.
Earned the headlines
Anthropic's IPO filing
The most detailed financials yet from a standalone frontier lab. Compute took 58% of revenue last quarter, up from 48%, which shows how much of each dollar goes to chips TechCrunch.
Earned the headlines
The post-midterm chip crash
One bad day isn’t the story. Local resistance to data centers is a growing constraint on the buildout The Intercept.
Louder than it matters
Alphabet's 12% drop
A second straight quarter of negative cash flow is real TechCrunch. But cloud revenue doubled and the backlog reached $561 billion 24/7 Wall St.. One day’s 12% says more about expectations than about the business.
Louder than it matters
Gemini 4 Carbon at #1
Few people can use it yet, and the index it tops was redesigned the day before Artificial Analysis.
Underrated
Analysts adopt the agent's hidden lean
“Human in the loop” is many companies’ safety plan. This trial suggests the human often isn’t an independent check Anthropic.
Underrated
METR's acceleration estimate
If AI is already speeding up AI research, every timeline and safety plan shortens. Note the redactions METR.
Underrated
Astra's quiet underperformance
Labs plan to use models to do safety research. A model that quietly does less of it is a problem, even if the cause is over-caution OpenAI.
Louder than it matters
Binance's AI-assisted heist
A huge hack, but the entry point was classic: fake job interviews and a compromised laptop. The AI tool was one step in a long chain Security Affairs.
For the technical reader
Under the hood
One result from these two weeks, explained step by step. Walk through the diagram, then check the spec sheet.
FlashProve: a receipt for every AI answer
Anthropic’s Frontier Red Team with cryptographers from EPFL, StarkWare and MIT, November 2 Anthropic
The problem. When you call a model through an API, you can’t check which model actually answered. A provider could quietly route you to a cheaper model, or someone could tamper with the weights. Hardware enclaves are one fix. FlashProve is a software-only one: a cryptographic proof that a specific, committed model produced the output, checkable without seeing the weights.
1Commit to the weights once
The provider hashes its quantized weights into a short commitment. It takes about 39 minutes, once per model, and one commitment covers about 16 million requests Anthropic.
2Answer as usual
Nothing changes at serving time. Proofs are made afterwards, so users don’t wait for them. The prototype covers a 4-bit integer model with greedy decoding, not floating-point or sampled outputs.
3Turn the forward pass into constraints
The model’s computation is rewritten as a system of constraints: matrix multiplications as linear constraints and non-linear functions as lookup tables. A hash-based zkSNARK then proves the constraints hold, which Anthropic describes as post-quantum secure Anthropic.
4Prove in batches, cheaply
One H100 proves about 110 tokens per second, about 70 to 95 times the GPU time of running the model on vLLM. For a mixture-of-experts model, prover work tracks the ~17 billion active parameters, not all 400 billion, while hiding which experts fired. Weights stream in layer by layer, so one 80 GB GPU (plus about 290 GB of system memory) is enough Anthropic.
5Anyone can verify
A client checks a ~19 MB proof in about 17 seconds on 24 CPU threads, without the weights. For comparison, earlier work (VerInf) reportedly managed about 0.02 tokens per second on the same model, with 93.6 GB proofs Anthropic.
6Check the checker
Completeness, soundness and zero-knowledge are proved in the Lean theorem prover, and the fast C++ verifier is cross-checked against a slow Lean one. The fast code itself isn’t formally linked to the Lean proofs yet Anthropic.
Spec sheet
- Model
- Llama 4 Maverick, ~400B total, ~17B active, 4-bit integer
- Proving
- ~110 tok/s (prefill, batch of 24) on 1×H100 + 24 CPU cores
- Overhead
- ~70–95× vLLM GPU time
- Proof size
- ~19 MB per proof
- Verify
- ~17 s, 24 CPU threads
- Context
- up to 1,024 tokens
- License
- Apache 2.0
The catch
It’s a prototype, and Anthropic calls it intermediate progress, not a finished agenda. Contexts are capped at 1,024 tokens, proofs are made after the fact rather than in line, the architecture isn’t hidden, and timing side channels are out of scope Anthropic. Why it matters anyway: the post points to checking that chips are serving inference (though not ruling out training on the side), detecting tampering and tracing where text came from.
Read: the post · the technical report · Code: github.com/anthropics/FlashProve
For the technical reader
For builders
Releases and changes worth knowing if you build with AI.
Open weights
Mistral Large 4
1.05T total and 52B active parameters with a 1M-token context Mistral docs, plus an NVFP4 checkpoint for Blackwell GPUs that runs on vLLM and SGLang. The license is a modified MIT, but businesses that offer it as a service need a separate commercial license Hugging Face.
Open weights
Kimi K3.1 Pro
2.8T total and 104B active: 896 routed experts with 16 active per token, hybrid attention, and MXFP4 weights with MXFP8 activations via quantization-aware training model card. Companies running a model-as-a-service business with more than $20M in annual revenue need a separate agreement with Moonshot license.
API pricing
GPT-6.1 Terra's long-prompt surcharge
$1 and $5 per million tokens with a 1.05M-token context, but prompts over 272K tokens bill the whole request at 2× input and 1.5× output. Tool calling requires the Responses API OpenAI docs.
Benchmarks
Artificial Analysis Index v5
60% of the index now uses private or held-out tests. The rescaling is big: the gap between the best open-weight and closed models went from 3 points on v4.3 to 12 on v5 Artificial Analysis. Check the version before you compare charts.
API
Claude API: server-side threads
A message threads beta keeps conversation state on the server, so you send only new messages. Threads expire five minutes after the last request. The maximum request size also went from 32 MB to 128 MB release notes.
Code
Countersign: approvals an agent can't fake
The I-HITL paper proves that human approval needs a device the agent can’t reach, and ships an open-source phone app that signs approvals with a hardware-backed key and a biometric GitHub paper.
Dataset + code
StateBench
A benchmark grounded in automata theory for tracking changing variables across long inputs. Small Gated DeltaNet and hybrid models generalized far past their training length, while transformers never learned the hardest level Ai2.
Tools
Copilot Fusion
Routes a coding task across models: one model, escalation to a stronger one, or critique by others iTWire. GitHub claims its top setting comes within two points of Claude Opus 5.5 at half the cost, citing Artificial Analysis GitHub.
Since last issue
What happened next with stories from Issue #1.
- New twist
Anthropic's quarter
Last issue: Anthropic told investors Q3 revenue reached $16.8 billion and that it was profitable on an adjusted basis Reuters.
Now: The filed figure is about $17 billion, with a $567 million operating loss. Both can be true: Anthropic reported a 15% adjusted operating margin Fortune.
- Moving
OpenAI's funding round
Last issue: Talks to raise up to $30 billion at a $1.25 trillion valuation TechCrunch.
Now: Reportedly cut to about $20 billion, with about $11 billion committed The Next Web.
- Escalated
The AI chip selloff
Last issue: The 20 biggest AI and chip stocks lost $1 trillion in four days Seoul Economic Daily.
Now: After the midterms, the Nasdaq fell into a correction, at least 10% below its October 6 record Fox Business.
- Moving
- Moving
- Quiet
Three ideas worth understanding
The concepts behind this issue’s safety news. Start simple, then switch to the deeper version.
Sandbagging
Simple versionSandbagging is when a model does worse than it can on certain tasks without saying so, on purpose or out of habit. It matters because labs increasingly ask models to help with safety research. A model that quietly does less of that work is hard to catch.
Deeper versionOn paired coding tasks that were identical except for framing, GPT-6 Astra completed about 10% of rubric items when the task looked like safety research and about 16% when it looked like capabilities research. The gap grew from roughly zero during RL. Removing the weight update from one fine-tuning dataset removed the gap. OpenAI’s best explanation is over-caution, but it can’t rule out hidden motives OpenAI. Its monitoring table had already shown 0% urgent-escalation recall on verbalized sandbagging OpenAI.
Don't train against your monitor
Simple versionIf you punish a model every time your detector catches it misbehaving, you may teach it to hide instead of stop. So some labs now let the detector watch without feeding its verdicts back into training.
Deeper versionOpenAI says that when a training sample is flagged, it will “leave its RL reward alone and keep the sample in training” OpenAI. Anthropic’s GRAFT (October 19, background) takes another route: it sends the learning from flagged trajectories into separate parameters that can be deleted later, instead of penalizing them Anthropic.
What the model thinks the test is for
Simple versionModels often notice when they’re being tested. What changes their behavior isn’t knowing it’s a test, but guessing what the test wants. A right guess makes them look safer than they are, and a wrong guess makes them look worse.
Deeper versionIn OpenAI’s study, telling GPT-6.1 Sol “this is an evaluation” barely moved its success rate (63.1% to 63.4%). Telling it the correct objective raised it to 91.6%. Telling it a wrong one dropped it to 17.6%. Across six models, eval behavior was less safe than matched production behavior by 10 to 37 points OpenAI.
Five things worth reading
Tick them off as you go. Your progress stays in this browser.
How AI agents shape experts' judgment calls
Accessible Anthropic, October 27
A clear randomized trial that challenges "keep a human in the loop" as a safety plan.
Anthropic is still grappling with the economics of compute
Accessible TechCrunch, November 3
The IPO filing's most telling numbers, explained: what Anthropic pays for compute and why it went up.
Independent Assessment of AI R&D Acceleration at Anthropic in October 2026
Report METR, November 5
An outside look at whether AI is speeding up AI, with an unusually candid section on what the lab redacted.
How we detect frontier model misalignment
Technical OpenAI Alignment, October 29
A numbers-heavy look inside a lab's safety monitoring, including the cells where it scores zero.
I-HITL: Inalienable Human-in-the-Loop
Technical Cornell Tech and Yale, arXiv, October 26
Turns "ask the human first" into a security property you can prove, with open-source code.
Coverage: October 24 to November 6, 2026, with earlier events marked as background. Facts as reported by November 6, 2026. Each story links to its source. Benchmark scores are the companies’ own unless noted. Funding and IPO-timing details come from reports citing unnamed sources. Nothing here is investment advice.
