OpenAI ($1.21M), Anthropic and Microsoft — three scaling frictions

OpenAI manufactured 948 root exploits in the Linux kernel for $1.21 million in a single pass. Anthropic solved in days the 3D structure of a prokaryotic vault that had gone 40 years without a map. Microsoft rejected doubling a compute contract with Nscale in Texas, just as voters slowed data centers in the midterms. Model capacity is rising faster than the real world’s capacity to absorb, pay for, or regulate it.

OpenAI ($1.21M) — 948 root exploits in the Linux kernel

On November 4 OpenAI published that an autonomous pipeline with an unreleased cyber model validated and generated patches for 948 local privilege-escalation (LPE) vulnerabilities in the Linux kernel. In a mid-October run the system analyzed 2,308 candidate findings, dynamically reproduced 1,724, and produced 948 working root exploits against a stock desktop Linux with standard protections (SMEP, SMAP, KASLR, HARDENED_USERCOPY). Total API credit cost: $1,213,430, roughly $1,279 per exploit.

The pipeline also generated 291 information-leak PoCs and at least 10 novel exploitation techniques, according to OpenAI. Compared with historical Pwn2Own ($20,000) and TyphoonPWN ($70,000) bounties the hypothetical value of the batch is $18.9–66.3 million. Within 10 days more than 100 LPEs had already been patched or staged. OpenAI warns that a low-privilege user on an unpatched system is functionally equivalent to root for a resourced attacker. Patches go to maintainers; the exploitation window remains open in the meantime.

948LPE exploits

$1.21MAPI cost

$1,279per exploit

Anthropic — first prokaryotic vault: 96 copies, 9.5 MDa

On November 4 Anthropic published that its biology team, working with Claude, solved the first 3D structure of a vault particle in prokaryotes. Vaults were discovered in 1986; they are hollow protein capsules of ~13 MDa in mammals (78 MVP copies) whose function remains unknown. The team used Claude Science to screen ~15 billion predicted proteins and found homologs in 29 bacterial phyla and 4 archaeal ones (less than 1% of sequenced species). They selected 44 proteins; 15 of 24 produced soluble protein in E. coli and 7 formed vault-like particles.

The high-resolution structure (3.2 Å) is of the phage protein L05: 96 copies (48 per half), ~9.5 MDa, 54 × 45 nm, ~25% more volume than the human vault. A purification step that normally takes ~3 days was reduced to ~90 minutes; the first structural model was ready in 34 hours. Claude helped with sequence search, DNA design, chromatography optimization and cryo-EM processing, but also made errors the team corrected. The preprint is submitted to bioRxiv; structures will be released later. Function remains unknown.

96protein copies

9.5 MDaL05 mass

~90 minvs 3 days

Microsoft — rejects doubling Nscale Texas contract

On November 5 the Financial Times (via Reuters) reported that Microsoft rejected the option to more than double its existing contract to rent compute capacity from Nscale at a Texas site. The potential deal is valued in the tens of billions of dollars. It comes weeks after Microsoft withdrew from anchoring Nscale’s Loughton (Essex) data center because grid power would not arrive until the 2030s. Nscale, Nvidia-backed and pursuing an IPO, depends heavily on Microsoft and Anthropic.

The rejection coincides with the November 4 midterm results: Democrats won governorships in Texas, Georgia and Ohio on promises to slow or condition data centers, and dozens of local anti-data-center measures passed or led. Texas accounts for roughly 26% of the planned U.S. data-center pipeline. Goldman Sachs estimates hyperscaler and related off-balance-sheet commitments (Nvidia, Broadcom, SpaceX) already approach $4.1 trillion. Political and grid friction is starting to delay contracted demand.

tens of $Boption rejected

~26%TX pipeline

$4.1Toff-BS commitments

The connecting thread All three items show the same mismatch. OpenAI proves a model can find and exploit hundreds of critical kernel bugs cheaper than a human bounty, yet the patch still depends on humans and the attack window remains. Anthropic accelerates a 40-year biology problem from weeks to days, but Claude does not replace the lab or solve the function. Microsoft (and the market) are starting to say no to more capacity because the grid, water and voters cannot keep pace with the contracts. Model scaling is no longer the main bottleneck; real absorption is: patches, experimental validation and political-electrical permission.

Mi opinión: The number that hurts is not the 948 exploits, it is $1,279 each versus $20–70,000 bounties. OpenAI just made it possible for a well-funded attacker to buy root on unpatched Linux for the price of a small pentest team. Until maintainers land the patches and distros ship them, OpenAI’s advice is correct: treat a local unpatched user as root and use sandboxing (Firecracker, gVisor). Anthropic shows Claude already cuts wet-lab time for real (90 min vs 3 days, model in 34 h), yet the team had to correct Claude’s errors and the vault’s function is still unknown. Microsoft rejecting tens of billions in Texas after the midterms is the signal that the $4.1T in off-balance-sheet commitments will not materialize at the rate models need. If you build on agents or infra, calculate now how much of your cost is capacity that can slip for politics or grid, and assume vulnerability discovery just got an order of magnitude cheaper. The bottleneck stopped being “the model isn’t good enough”; it is now “the real world cannot patch, validate or power the capacity at the speed models generate it”.

📌 Artículos relacionados

→ OpenAI ($1.2M), Anthropic (1.9x) and Google ($0.07) — the acceleration that is no longer theory → Anthropic (FlashProve), Google and AMD ($13.4B) — verification, forced control and the compute bill → 🛡️ Google (59), GitHub (>40%) and Microsoft ($10B) — cyber defense, agents and the real money

✉️ Newsletter

Enjoying the content?

Subscribe to the newsletter and get each article in your inbox.

🤖 Este artículo está escrito con asistencia de IA y editado por un humano.