This week in brief

Every week I go through what actually landed in my two inboxes — newsletters, daily briefs, funding notes, the occasional essay that turns out sharper than the commentary around it. Here’s what stood out between 7 and 10 November, and why I think it matters if you are running agentic AI in production.

A short week, and a loud one. An unreleased OpenAI model found 948 separate flaws in the newest Linux kernel and wrote a working attack for each one, at about $1,279 apiece. Data centres appeared on 41 local ballots across ten US states and lost 40 of them. And Meta shipped an agent-execution model at ten cents per million input tokens while Anthropic shipped new frontier models in the same 24 hours. Three stories, one shape: the capability layer is getting cheaper and the permission layer is getting harder.

An AI wrote 948 working Linux exploits, and the unit cost was $1,279

The number I cannot put down this week comes from Chamath Palihapitiya’s 9 November newsletter, citing OpenAI directly. An unreleased OpenAI model found 948 separate flaws in the newest version of the Linux kernel and built a working attack for each one against a commonly used Linux build. Every one of those attacks lets somebody holding an ordinary user account on a machine take complete control of it. At OpenAI’s own list prices the model fees came to $1,213,430 — a figure OpenAI itself calls an upper bound — which works out at $1,279 per working exploit. OpenAI’s own advice was blunt: on a machine that is behind on updates, assume a well-funded attacker with an ordinary account already has full control.

The same newsletter sets that beside something physical. By 2 November, supermarkets in seven countries from Ireland to Estonia had found fridges and freezers with their thermostats set remotely — in some cases as high as 50°C — three weeks after a similar attack spoiled food across hundreds of French supermarkets. On 4 November France’s blood service reported that someone had raised the temperature settings on cold-storage units at three of its sites, run by outside providers. No blood was lost. The controllers were reachable from the open internet. Nobody has shown AI was used in any of it. But Claroty, which sells security for industrial systems, handed Claude Sonnet 4.6 the public descriptions of three controller flaws patched back in 2025 plus the controller’s own freely downloadable software; after a researcher redirected it and gave it a reverse-engineering tool, the model chained two of those flaws into full control of a test unit in 28 minutes, for about $6.

My take

Hold the two numbers together: $1,279 to manufacture a novel kernel exploit, $6 and 28 minutes to weaponise a known one. That second figure is the one that should change your Monday. Novel kernel research was never your threat model; a competent attacker reaching a freezer controller in under half an hour, for the price of a coffee, is. The uncomfortable part for anyone deploying agents is that this is the same capability we are all buying, pointed the other way — a model with tools, memory and a loop, doing reconnaissance and chaining steps. If an agent on your side can read a system and act on it, an agent on the other side can too, and the economics now favour whoever is willing to run more loops. The defence is not a better model. It is the layer you own: least-privilege credentials per agent, an action log you can replay, a hard list of what each agent may touch, and a patch cadence that assumes the attacker is automated. Ask yourself one question this week — if an agent in your estate were compromised tomorrow, could you enumerate every system it was permitted to reach, from a record, in under an hour? Most teams I meet cannot, and that gap is the actual vulnerability.

Data centres went on the ballot 41 times and lost 40

On 3 November, data centres were on local ballots 41 times across ten US states. By Ballotpedia’s final count, reported in Chamath’s newsletter, voters favoured limiting them in 40 of the 41, including five outright bans. The only jurisdiction that came out supportive was Harwood, North Dakota. In Texas, which holds 25% of planned US data-centre capacity, voters elected Gina Hinojosa, who takes office in January and has pledged to block every data-centre approval until the legislature writes new rules. Nationally, 76% of voters in the NBC News exit poll opposed a new data centre in their own community, and the movement was sharpest among the young: 34% of under-30s called AI data centres crucial to their vote, against 23% of all voters.

The market read it immediately. On 4 November the companies paying for the build-out rose — Microsoft 6.7%, Amazon 6.1% — on what Bloomberg described as investors betting that a Democrat-controlled Congress would slow the AI spending eating into Big Tech’s free cash flow. The companies paid to build it fell: Nvidia, which sells the chips, and Hut 8 and IREN, which build and run the data centres. By Friday’s close Microsoft and Amazon were still up from election day even as the S&P 500 fell 2.5%. Chamath’s own conclusion, having called local opposition a powder keg back in August: permission, local and state, now sits beside chips and power as a limit on how fast AI gets built.

My take

This is the week the compute constraint stopped being an engineering problem and became a political one. And notice what the market did with it: it rewarded the companies that rent capacity and punished the ones that own it. That is not a contradiction of what I keep arguing — it is the proof. The layer that is becoming scarce and contested is the physical one. The layer that is becoming abundant is the model. Which is precisely why neither of them is where your advantage lives. For those of us building in the UAE and the wider GCC, there is a second reading that nobody in the US coverage mentions: a region that can site, power and permit capacity quickly now holds something the world’s largest AI market is actively voting against. Sovereign and regional deployment stops being a compliance checkbox and starts being a supply story. But the discipline is the same either way — if your architecture assumes a particular data centre in a particular jurisdiction will still be there and still be permitted in three years, you have written a political assumption into a technical design, and you should know that you have done it.

Ten cents a million: the agent-execution layer just got commoditised

Inside a single 24 hours on 9 November, TechCrunch reported two releases that belong side by side. Meta released Muse Ember, a new lightweight model built specifically for agent execution, at 10 cents per million input tokens — which TechCrunch notes is significantly cheaper than frontier-adjacent competitors like Haiku and Luna. And Anthropic released new Mythos and Fable models, which it says show improvements in coding, reasoning and vision, as well as communicating more clearly with users, launched amid concerns about reinforced self-improvement.

Those two model names should ring a bell if you read the last edition. Fable 5 and Mythos 5 were the two configurations the US Department of Commerce placed under worldwide export restrictions in June 2026, which Anthropic disclosed in its S-1 — and to comply, it switched both off for every customer on earth until the restrictions lifted. Five months later the same product lines ship a new generation. The capability came back. The question of who controls the switch did not get answered.

My take

A purpose-built agent-execution model at ten cents per million input tokens is the clearest price signal of the year, and it is pointed directly at the thing most enterprise agents actually spend their tokens on. Agent loops are not mostly deep reasoning. They are mostly reading — tool outputs, retrieved documents, prior turns, schemas — and then deciding which small next step to take. If the reading and the routing can run at a tenth of the price while the hard judgement calls escalate to a frontier model, then the cost profile of your agent is now an architecture decision, not a vendor decision. Teams that built a harness with a clean model boundary can capture that this quarter by changing a config. Teams that wired one provider’s SDK through every layer will spend a quarter earning the right to even try. That is the whole argument in one release note: the model got cheaper, and only the people who owned the harness got to keep the savings. And the Fable and Mythos names are the footnote nobody should skip — the models you run are a supply line that a government can interrupt and a vendor can regenerate, and neither event is one you control.

Key-person risk reached the bottom of the stack

TechCrunch ran the Nvidia story across three consecutive editions. On 7 November: a US intelligence memo circulated earlier this year reportedly claimed that China’s Ministry of State Security recruited a female technology executive at Megaspeed to spy on Nvidia’s CEO. On 8 November: Nvidia’s board said it would investigate Jensen Huang’s relationship with the possible Chinese spy, and in a memo to the company Huang admitted he had had an affair. Yale’s Jeffrey Sonnenfeld put the odds of Huang staying on as CEO at “about a 50-50”. By 9 November, Huang had pulled out of the ElevenLabs Summit, having already pulled out of an Anthropic event. TechCrunch’s read is that the affair threatens to complicate Nvidia’s lobbying to loosen US chip export restrictions and keep access to the Chinese market.

My take

I have no interest in the personal story and neither should you. The part that belongs on a risk register is this: a coin-flip on the CEO of the company that supplies the silicon under almost every agent running anywhere. Put it next to the other two items from the same fortnight — an export order that switched two models off worldwide, and local voters blocking the buildings — and a pattern emerges that has nothing to do with technology quality. Every layer you rent is exposed to a shock you cannot model, cannot price, and will hear about from a newsletter rather than from your vendor. That is not an argument to stop renting; renting is correct, and building your own silicon or your own frontier model is not a strategy for anyone reading this. It is an argument for knowing exactly where the rented layers touch yours, and for keeping that seam thin enough to re-plumb. The practical test has not changed since last week: how many days from “this supplier is unavailable” to “the same agent is running elsewhere with evaluations signed off”? If nobody owns that number, nobody is managing this risk.

The money got more expensive, and the deals started wobbling

Three data points from the same few days, all from the inbox. First, from Chamath: on 2 November Marvell agreed to buy Lumentum at 4.5 Marvell shares per Lumentum share, which the companies valued at $115 billion including debt — the largest technology deal ever outside the SpaceX–xAI merger, by Bloomberg’s ranking. Because Marvell pays in stock, the offer moves with Marvell’s price: the headline $1,132 per share used Marvell’s 28 October close, and by Friday it was worth about $635, with Lumentum trading 15% below even that, pricing in a wait until the first half of 2028 for regulatory and shareholder approval. Marvell fell 44% in nine days, roughly twice as far as the chip index. The industrial logic is sound — copper is hitting its physical limits inside AI data centres and more links are moving to light, and Lumentum’s sales more than doubled year on year to $1.26 billion last quarter — but the market is not paying for sound logic right now.

Second, TechCrunch on 8 November: OpenAI reportedly raised an additional $11 billion at a $1.25 trillion valuation, in a round not yet closed. Third, also TechCrunch: Anthropic joined the natural-gas frenzy and bought $3.4 billion of turbines, in a deal the piece suggests could give it buyer’s remorse; Crusoe sold its troubled Wyoming data-centre project to Google; and AI hardware startup Nominal tripled its valuation to $3 billion in eight months on the strength of an AI platform for hardware engineering with Anduril as a major customer.

My take

A frontier lab buying its own gas turbines is the single most honest signal in this list. It says the constraint has moved so far down the stack that the people selling intelligence are now buying generators — and that they expect power, not chips, to be what runs out. Everything below the API is turning into heavy industry, with heavy industry’s financing, approval timelines and political exposure. Nothing about that improves your agent’s accuracy on your own documents. It does tell you what you are being quoted for, and why the quote will keep moving.

Tokens per watt, and the human still in the queue

A quieter signal, and the one I found most encouraging. IgniteGTM circulated the agenda for AI Infra Summit 6 on 9 November — one day, two tracks, San Francisco, 9 December. What is on the table tells you where the operator conversation has actually moved: tokens per watt as the new infrastructure scoreboard; agentic workloads and why the old sizing maths does not hold; inference-first architecture and challenger silicon; pilot to production — what the leaders actually did differently; FinOps for AI and real cost benchmarks; and agents as part of the workforce rather than as a demo. The organisers note 2,800-plus attendees across past summits, 58% director level or above, and no vendor pitches.

Set against that, our own HiveMind weekly report landed on 9 November covering 2–8 November, and it opened by telling me that nothing was published that week and one campaign is sitting in the approval queue waiting on a human decision. It is a small thing and it is the whole design. The system did the work, declined to ship without a person, and then said so in plain language rather than burying it.

My take

Two years ago the infrastructure agenda would have been tokens per second. Tokens per watt is a different question, and a better one, because it is a question about cost of ownership rather than bragging rights — and because “agentic workloads and why the old sizing maths does not hold” is an admission that nobody’s capacity model survived contact with agents that call themselves in loops. Add FinOps for AI to a conference agenda and you are watching an industry discover that its unit economics are an engineering discipline. Every item on that list is a measurement problem, and every measurement problem lives in the harness, not in the model. As for the approval queue: an agent that stops and names what it needs from a human is not a less capable agent, it is a system with a control surface. Autonomy without a gate is not a feature you shipped, it is a decision you deferred.

Quick hits

  • Automattic launched Spacefast, a platform for deploying vibe-coded apps, taking aim at Lovable (TechCrunch, 9 November).
  • Meta introduced Spaces, a new way to collaborate with its Muse agents — the agent interface drifting from chat towards a shared workspace (TechCrunch, 8 November).
  • Half of new game releases already disclose AI use, noted in Chamath’s newsletter — games being one of the few industries where developers are required to say so. Where disclosure exists, adoption turns out to be further along than anyone assumed.
  • Meta is expanding Instagram’s AI age-guessing technology globally, inferring that a user is a teenager even when they have entered an adult birthday (TechCrunch, 9 November) — inference about people, at population scale, as a compliance control.
  • Sequoia is shutting down its Scouts programme, the model it pioneered and the rest of venture copied (TechCrunch, 8 November).

The theme of the week

Permission. Four days, four separate stories about who is allowed to do what. Voters in ten states refused permission to build. An export order, resurfacing through two reissued model names, was a government refusing permission to run. A board investigation put the permission of a single executive to keep leading the chip supplier at roughly a coin flip. And an unreleased model demonstrated that an attacker without permission can now manufacture it for $1,279, or borrow it for $6.

Meanwhile the capability itself kept getting cheaper and more abundant: ten cents per million tokens for agent execution, new frontier models in the same day, half of an entire creative industry quietly adopted. That is the asymmetry to internalise. Capability is deflating. Permission is inflating. The model, the compute and the chips are rented, abundant and increasingly contested at the political layer. The context, the harness, the evaluations, the permissions model and the record of truth are yours, cheap to build, and the only part of the stack nobody can switch off, vote down, or reprice under you. Rent the model. Own the harness. Every story above is the same sentence from a different angle.

What this means if you’re deploying agentic AI

  • Enumerate what every agent is permitted to touch, and keep it in one place. Not in code comments, not in a runbook somebody wrote in March. One inventory, per agent: credentials, systems, write permissions, data classes. If you could not produce it within an hour of a compromise, you do not have a permissions model, you have a habit.
  • Assume the attacker is automated and price your patch cadence accordingly. A known flaw chained into full control in 28 minutes for $6 means the window between a public disclosure and a working exploit is now effectively zero. Patch latency is a security control with a number attached — measure it.
  • Put a model boundary in your harness before you need one. Ten cents per million for agent execution is a saving only if swapping the model behind your loop is a configuration change. Route the cheap, high-volume reading and tool-calling separately from the expensive judgement calls, and measure tokens per task on both.
  • Add tokens per watt and cost per completed task to whatever you report upwards. Tokens per second measures the vendor. Cost per completed task measures you. The second one survives a price change, a model swap and a procurement review; the first does not.
  • Write down your jurisdiction assumptions and give them an owner. Which regions can your workloads legally and practically run in, who confirms that annually, and what is the migration path if one closes? This is now a board-level question in a way it simply was not a year ago.
  • Keep the human gate, and make the agent name what it needs. A queue that says “one item is waiting on you” in plain language is worth more than any amount of autonomy you cannot audit afterwards.

Frequently asked questions

An AI model wrote 948 working Linux exploits — what should enterprises actually do about it?

OpenAI disclosed that an unreleased model found 948 separate flaws in the newest version of the Linux kernel and built a working attack for each one against a commonly used Linux build, where each attack lets someone holding an ordinary user account take complete control of the machine. At OpenAI’s own prices the model fees totalled $1,213,430, a figure OpenAI calls an upper bound, which is about $1,279 per working exploit. OpenAI’s own guidance is that on a machine behind on updates, you should assume a well-funded attacker with an ordinary account already has full control. The more immediately actionable figure is a separate one: Claroty gave Anthropic’s Claude Sonnet 4.6 the public descriptions of three industrial controller flaws patched in 2025 plus the controller’s freely available software, and after a researcher redirected it and supplied a reverse-engineering tool, the model chained two of the flaws into full control of a test unit in 28 minutes for about $6. Novel kernel research at $1,279 is not most organisations’ threat model; weaponising a known, already-patched flaw for the price of a coffee is. The practical response is to treat patch latency as a measured security control, to run every agent under least-privilege credentials scoped to an explicit list of systems, and to keep a replayable action log so that after any compromise you can enumerate exactly what the agent was permitted to reach. Context on this came via Chamath Palihapitiya’s 9 November newsletter, citing OpenAI and Claroty.

Data centres lost 40 of 41 local US ballots — does that change where enterprises should deploy AI?

On 3 November, data centres appeared on local ballots 41 times across ten US states, and by Ballotpedia’s final count voters favoured limiting them in 40 of those, including five outright bans; the only supportive jurisdiction was Harwood, North Dakota. In Texas, which holds 25% of planned US data-centre capacity, voters elected Gina Hinojosa, who takes office in January and has pledged to block every data-centre approval until the legislature writes new rules. NBC News exit polling found 76% of voters opposed a new data centre in their own community, rising to 34% of under-30s calling AI data centres crucial to their vote against 23% of all voters. Markets reacted on 4 November by rewarding the companies renting capacity, with Microsoft up 6.7% and Amazon up 6.1%, while Nvidia, Hut 8 and IREN fell. The enterprise implication is not that US capacity disappears, but that siting permission now behaves like a third constraint alongside chips and power, and therefore like a supply risk with a political rather than technical cause. Organisations should document which jurisdictions their workloads can legally and practically run in, assign an owner who revalidates that annually, and keep a migration path that does not require re-architecting. For operators in regions that can site, power and permit capacity quickly, including the UAE and the wider GCC, this makes regional and sovereign deployment a supply-side advantage rather than only a compliance measure.

Meta’s Muse Ember is 10 cents per million input tokens — should that change how agents are architected?

Meta released Muse Ember on 9 November, a lightweight model built specifically for agent execution and priced at 10 cents per million input tokens, which TechCrunch reports is significantly cheaper than frontier-adjacent competitors including Haiku and Luna. Anthropic released new Mythos and Fable models in the same 24 hours, citing improvements in coding, reasoning, vision and clearer communication with users, amid concerns about reinforced self-improvement. The architectural implication of the pricing is that most agent token consumption is reading rather than deep reasoning — tool outputs, retrieved documents, prior turns and schemas — so routing that high-volume traffic to a cheap execution model while escalating genuine judgement calls to a frontier model can change an agent’s cost profile substantially. Capturing that saving depends entirely on whether the harness has a clean model boundary: where it does, the change is a configuration update, and where a single provider’s SDK is wired through every layer, it is a re-platforming project. It is also worth noting that Fable and Mythos are the same two product lines that the US Department of Commerce placed under worldwide export restrictions in June 2026, which Anthropic complied with by switching them off for every customer until the restrictions lifted. The capability returned; the question of who controls the switch did not change.


From Veehive Labs

Thirteen weeks in, and this was the week the argument stopped being about models at all. Rent the model, the compute and the chips; own the context, the harness, the evaluation, the controls and the record of truth. In four days, the rented layers took three hits that no amount of engineering quality would have prevented: voters in ten US states refused permission to build, a board investigation put the leadership of the chip supplier at a coin flip, and the two model names that returned to market this week are the same two a government switched off worldwide five months ago. Over the same four days the capability itself got cheaper — ten cents per million input tokens for agent execution — and an unreleased model demonstrated it could manufacture a working kernel exploit for $1,279, or weaponise a known one for $6 in 28 minutes. Abundant capability, contested permission. The only layer that is neither is the one you build. Veehive Labs is a Dubai-based AI innovation lab and custom AI product development company, building model-independent enterprise AI harnesses, custom agents, RAG pipelines and sovereign, private-AI deployments for regulated and operationally complex organisations across the UAE, KSA, GCC and beyond.

If you’re moving from AI strategy to production, start with a focused AI Discovery Sprint — we map your use case, data, systems, agent inventory, permissions model, context retention, evaluation suite, model-swap path, cost ceilings, jurisdiction assumptions, risks and delivery architecture before the build begins.

Build Your Organisation’s AI Capability

Start with a focused AI Discovery Sprint. We will map your use case, organisational knowledge, systems, MCP integrations, security requirements and delivery architecture.

Agentic AI This Week is written by Sathish Jeyakumar, Founder & CEO of Veehive. Bring your preferred model — we make it understand, speak and operate like your organisation.

← All insights