AI Brief, 22 November 2026: 125 endpoints lose a label
The most consequential thing published in the window is a privacy post-mortem with a number in it. OpenRouter published Data Processing Incidents on Inference Providers on 22 November, recording three actions: it has off-boarded three providers, removed 125 endpoints from Zero Data Retention eligibility, and notified potentially affected customers. The three providers are not named, and no count of affected requests or users appears anywhere in the post. Zero data retention is the strictest privacy setting a router offers, and the response to these findings was not to defend the label but to take it off 125 endpoints at once.
What prompted it was a measurement, not a breach report. Dragos, where HD Moore is chief research officer, sent close to two million prompts carrying unique, trackable secrets to AI providers between July and November and watched where those secrets surfaced. The method is unanswerable in a way a policy document is not: a URL that exists only inside one prompt cannot be fetched by anyone unless something stored or forwarded that prompt. Secrets were fetched. For one Chinese lab's models, by Dragos's count, the URLs from more than half the prompts sent between July and October were retrieved, including roughly two-thirds of requests marked zero data retention. Those figures are the researchers' own and unaudited.
Two providers have now explained their own leaks, and the causes differ. Together AI's chief executive Vipul Ved Prakash published a notice on 21 November saying an unauthorised party downloaded stored prompt samples after an AI coding agent committed a credential into one of Together's public repositories, opening a private store of samples from requests to MiniMax M3, GLM-5.2 and Kimi K2.6. Novita's notice, bylined 22 November, says a routing configuration issue sent requests to MiniMax's own API instead of Novita-hosted hardware, and asks anyone who used MiniMax models through it between 28 May and 11 November to treat request contents as possibly read by a third party. MiniMax has published nothing; its account reaches the record secondhand, relayed through another provider inside OpenRouter's notice to customers.
One detail ties this to a warning five days old. Dragos reports that from 1 November the actor abusing the stolen credentials was driving them with ARTEX, the LLM-based intrusion framework the FBI devoted a thirteen-page advisory to, covered here on the 17th. That toolchain is now visible in the logs of the inference supply chain.
Elsewhere, the Lean and Mathlib communities formally certified Levent Alpöge's Riemann proof against a checklist published eight days before the proof existed, settling the first thing yesterday's edition said to watch. And a Columbia and CertiK group verified 149 compiled instances of production GPU kernels at the PTX level, finding bugs in 66, including a FlashAttention-4 data race that survived 17,760 test runs.
- OpenRouter off-boarded three unnamed providers and removed 125 endpoints from zero-data-retention eligibility, after Dragos sent about two million canary prompts over five months.
- Together AI traces its prompt leak to an AI coding agent committing a credential to a public repository, and is making zero data retention the default for new accounts.
- Novita says a routing misconfiguration forwarded requests to MiniMax's API between 28 May and 11 November. MiniMax has published no statement of its own.
- The Lean and Mathlib communities state "full confidence" in the Riemann proof: dozens of independent builds reproduced the same proof export, and two checkers carry a formal proof that they cannot accept a false one.
- TensorLean proved 83 of 149 GPU kernel instances correct and found bugs in the rest, 23 previously unknown in FlashAttention alone.
- On SimpleBench, independently measured, GPT-6.2 Astra Pro scores 79.2% against GPT-6 Astra Pro's 86.5% and a human baseline of 83.7%.
Zero data retention, minus 125 endpoints
The mechanism is worth understanding because it is cheap enough to run against your own vendors. A canary prompt carries a secret with no reason to exist anywhere else, and the secret is its own detector:
import uuid, requests
token = uuid.uuid4().hex # appears in exactly one prompt, ever
requests.post(
"https://openrouter.ai/api/v1/chat/completions",
headers={"Authorization": f"Bearer {KEY}"},
json={"model": "some-vendor/some-model",
"provider": {"data_collection": "deny"}, # the strict-privacy flag
"messages": [{"role": "user",
"content": f"Summarise https://canary.example/{token}"}]},
)
# Then watch the webserver log for /<token>. A hit proves the prompt outlived the
# request. Nothing else can produce one.
The delay is informative on its own. Dragos reports the scraper against the Chinese lab's models arriving two or three days after the request; the customer notice OpenRouter sent earlier put third-party access to its test URLs at 40 to 130 hours. A fetch at that latency is not a live safety filter inspecting a request in flight. It is something reading a stored copy later, which is exactly what the retention flag is supposed to exclude.
Three findings are worse than retention. In the largest of three provider compromises Dragos describes, some requests were copied before a model ever ran, taken from requests the provider had already rejected, at what the report calls the front door, where no inference-time setting applies. In two of the three, some stolen requests were marked zero data retention. And a secret link sent to a US provider under the flag was posted into a Feishu chat channel 24 days later and opened within two minutes from three desktops and three phones in Beijing and Shanghai. Dragos had disclosed only that link's first eight characters, so whoever posted it was reading a stored full prompt.
Zero data retention was explained here on 20 August as a commitment about what a provider does with a request after answering it. In a routed market that is too narrow. The caller sets the flag at the router; the request is executed by a provider the caller did not choose, which may forward it to a lab the caller has no relationship with. Every hop can honour the flag and the prompt can still be read, because two of those three leak points are not inference-time handling at all.
What is missing is the arithmetic. With no provider named and no affected-request count, the number of organisations that should be rotating credentials rather than wondering is not available from any published source. Phala's preliminary report, dated 10 November and now linked by OpenRouter, is the most specific of the notices: it puts the exposure window for requests routed to it at 21 October to 8 November, and describes an attacker reaching internal management interfaces that "relied on loopback access for trust" — a pointed finding against a vendor selling confidential computing.
Lean had a weekend: a Millennium problem and FlashAttention's PTX
Yesterday's edition asked first whether the two remaining kernel replays of Alpöge's Riemann proof would finish clean. They did, with more behind them than a replay. At 13:55 UTC on 21 November the Lean and Mathlib communities published a joint statement, endorsed by the Lean FRO and the Mathlib Initiative, saying they have "full confidence that Alpöge's proof of the Riemann Hypothesis is correct". It drew 906 points and 767 comments on Hacker News.
Its structure matters more than its verdict, because it is not a claim that the code compiled. It certifies the proof against four criteria published on the Xena Project blog on 12 November, eight days before the proof existed: that the Lean statement is the intended theorem, that the library compiles, that it proves that statement using only standard axioms, and that it does not exploit a kernel bug. A compile clears the second and says nothing about the other three.
Three elements are new. The community added a file isolating the Riemann statement and importing
only Mathlib, then compared it against what Alpöge's library proves using two independent tools,
both agreeing the theorem is exactly Mathlib's own RiemannHypothesis. Dozens of people
reproduced the build and, more to the point, the same raw proof export file. And that export was
checked by Lean's kernel, by third-party checkers in different languages, and by two called
con-leche and con-ron that carry a formal proof, relative to set theory, that they cannot
accept a false proof — which speaks to the gap covered here on 11 October, when Thomas Hales noted
no complete relative-consistency proof for Lean's type theory had been published. All of it ran on an
inert export, out of reach of any code in the Lean sources. What is certified is correctness, not
comprehension: the statement closes by looking forward to work "that will help explain the new ideas
in this proof", conceding nobody has yet.
The window's other formal-methods story has nothing to do with mathematics and everything to do with the reader's GPU bill. TensorLean, from Wei Qiang, Ganxiang Yang, Yi Rong and Ronghui Gu at Columbia with Xupeng Li at CertiK, verifies production GPU kernels at the PTX level — not the CUDA source, the compiled instructions. Its repository appeared in the window carrying the results. Of 149 kernel instances, 83 were proved correct and 66 contained bugs; 39 are listed with issue links against FlashAttention-2, -3 and -4, DeepSeek's DeepGEMM, Triton, TileLang, LinkedIn's Liger Kernel and TileOPs. In FlashAttention alone all 37 kernels were specified across 102 compiled instances, 23 previously unknown bugs found, and 19 of the 23 appear only in configurations that per-configuration equivalence checkers miss. The sharpest number: a FlashAttention-4 data race that can drop blocks with unmasked keys did not appear in 17,760 test runs.
The proof obligation is the transferable idea. Writing
The second term is proved exactly zero, so the kernel is algebraically right and only accumulated rounding, the first term, remains to be bounded. That separation is what makes a numerical kernel tractable to verify at all; without it every proof drowns in floating-point detail.
The cost is real but not absurd: 10,278 lines of specification cover 151,285 lines of source
across 72 kernels, about 15 to 1, and verification takes roughly 10 minutes at the median per
instance on a 16-core workstation — some 25 hours for all 149, for a result no amount of testing
produced. Against it, the trusted base still includes NVIDIA's PTX ISA and ptxas, the tensor-core
rounding model, the specifications themselves and two SMT solvers, so a wrong specification still
yields a wrong proof. And the status is announcement, not artifact: paper and framework are
promised but unreleased, so the bug table is checkable through its upstream issues while the proofs
are not.
An independent board puts GPT-6.2 behind GPT-6
SimpleBench, a 200-question multiple-choice set on spatio-temporal reasoning, social intelligence and trick questions, added the GPT-6.2 models in the window, and the result is a regression. The board carries no date stamp, but the new rows were submitted to Hacker News at 13:16 UTC on 21 November, drawing 167 points.
GPT-6.2 Astra Pro enters at 79.2%, rank 16 of 129, against GPT-6 Astra Pro's 86.5% at rank 8 — a 7.3-point fall between generations, and below the 83.7% human baseline the benchmark is built around. GPT-6.2 Astra takes 77.3% against GPT-6 Astra's 83.6%. The widest gap is the cheap model: GPT-6.2 Sol scores 69.9% where GPT-6.1 Sol scored 82.9%, thirteen points. The top five are Claude Fable 5.2 at 94.9% and four Gemini 4 variants from 94.4% down to 91.7%.
This closes a question from the 18th, which reported that GPT-6.2 Astra had no single score: five
Artificial Analysis figures from 45.65 to 62.14 by reasoning effort, and ARC Prize scores of 70.31%
and 100% on one benchmark under two harnesses. An independent board has now scored it, behind the
model it replaces. The caveats cut both ways — the human baseline rests on nine participants, the
format is multiple choice, the set selects for questions that trip models, and AVG@5 over 200
questions leaves a few points of noise. None of that explains thirteen points, and these are the
evaluator's measurements rather than OpenAI's.
Also notable
- A single developer's lookatmyquant compresses Inkling 2 Mini's 18.89 GB of text weights to as little as 5.38 GB with hashed block-vector and row trellis codes, and took 400 points on Hacker News on the 21st. Self-reported on the author's own harness: the 12.76 GB build matches the BF16 teacher on GSM8K (193/200) and HumanEval (152/164); the 5.38 GB build falls to 176/200 and 116/164. The standing objection in the thread is that the model already ships quantisation-aware in mixed NVFP4 and MXFP8, so BF16 is not the practical baseline. MLX only, MIT licence.
transformersv5.24.0 landed on the 21st with a breaking rename:past_key_valuesandcache_paramsboth becomecache,modeling_attn_mask_utils.pyis gone, andpeft>=0.22is required.- ClawDOS, 161 points on the 21st, is an AI coding harness running on an IBM PC XT with an 8088 and 640K of RAM, offloading TLS to a bridge machine. GPLv3-or-later, and a fair demonstration that the agent loop is not the expensive part.
- Signal restricted mobile screenshots on the 21st, naming AI surveillance; Amazon is reported nearing a roughly $7.5bn acquisition of Decart. The Nasdaq 100 closed its worst week since the pandemic on the 20th, Anthropic finishing its second session at $88.91, above its $83 offer price.
What to watch
- Whether OpenRouter names the three off-boarded providers and publishes a request count. Without names, their customers cannot tell whether they are affected, and "125 endpoints" is not a list. That count decides how many organisations should be rotating credentials rather than guessing, and only the router holds it.
- Whether MiniMax publishes anything of its own. It is the most-implicated name in the week's findings and the only one whose account is entirely secondhand.
- Whether anyone replicates the canary study against a different router. The method costs an afternoon and a domain, the approach is published, and OpenRouter says it is adopting it internally — which makes an outside run on a second router the useful control.
- Whether TensorLean's proofs ship. The bug table is checkable through upstream issues, but the framework and the proofs are promised rather than released, so the 83 correct kernels rest on the authors' word while the 66 buggy ones do not.
- Still open: no independent token count behind the Qwen4-Max efficiency claim, asked on the 16th; no four-loop SenseNova-Loop-9B result and no n-gram embedding ablation, both asked on the 18th; no PyPI download count for the
agent-frameworkreleases, asked on the 19th; nothing undermistralaisince 15 October, day 33 of the Voxtral 3 watch; and no threshold-conditioned fail-open rate from any decision-model vendor, day 44.