Independent/Reader-funded/Infrastructure, not tokens
DeAINEWS

AI you control — open models, private inference, and the networks that run them.

Provider Policy & Trust

The 9 DeAI stories that mattered this week

The week open-source inference itself got breached: NEAR AI's TEE didn't stop an SGLang RCE, xAI watermarked the Grok API by default, and Nvidia lost its CEO.

DeAI is powered by Morpheus (mor.org). We cover competing providers on the same terms — see our methodology.

A single GPU inference server rack with its front panel door ajar in a dim datacenter corridor, circuit boards exposed under one amber status light, depicting the week's lead story: the SGLang remote-code-execution breach that reached inside NEAR AI's inference stack. Illustration: DeAI
A single GPU inference server rack with its front panel door ajar in a dim datacenter corridor, circuit boards exposed under one amber status light, depicting the week's lead story: the SGLang remote-code-execution breach that reached inside NEAR AI's inference stack. Illustration: DeAI

The week the attack surface introduced itself. NEAR AI — a provider whose entire pitch is verifiable private inference — paused its Cloud service on November 9 after unauthorized access to internal systems, and the incident report that followed on the same day delivered the uncomfortable detail: an attacker exploited a remote-code-execution vulnerability in SGLang, the open-source inference engine, and the confidential-computing enclave did not stop it. About 100 accounts exposed, 18 agents' data reached, two customers' API keys found in logs and revoked. The report's own admission is the week's lesson: hardware isolation protects a prompt from the operator and the host, not from a bug in the software running inside the enclave.

Around that spine, the week split into three currents. The policy floor rose and the chip throne emptied: xAI turned on invisible watermarking by default across the Grok API twenty-five days before the EU AI Act's marking deadline, and Nvidia announced on November 12 that Jensen Huang — the CEO whose export controls and Hugging Face acquisition defined the hardware layer open weights run on — resigned effective immediately, with CFO Colette Kress interim. The proof economy argued with itself: a viral checkpoint-forensics report claimed six Chinese open-weights models carry Blackwell fingerprints without publishing code, while Nous Research released Arete-27B with weights, eval, and training recipe public, and the tracker logged its first positive check evidence in seven cycles — still zero verified. Nine stories mattered; here is what each one changes.

Key facts

  • ~100 accounts / 18 agents — what NEAR AI's incident report says an SGLang RCE flaw exposed, on a legacy non-TEE deployment, with two customers' API keys found in logs and revoked; the report concedes confidential computing does not protect against a vulnerability in the software inside the enclave.
  • 7 model variants — Grok models watermarked by default on every xAI API endpoint since November 7, per xAI's own developer documentation, with a per-model team-admin opt-out that is audit-logged; the EU AI Act's Article 50(2) marking deadline lands December 2.
  • 33 years / effective immediately — Jensen Huang's run at Nvidia, ended by a board announcement on November 12 with no reason given; CFO Colette Kress is interim CEO, and the $12.93 billion Hugging Face acquisition announced in September has not been rescinded.
  • 6 models / 0 reproductions — the Blackwell checkpoint-forensics report names DeepSeek-V4, Kimi K3, Kimi K3.1-Pro, MiniMax-M3, and Ling-3.1 variants as carrying Blackwell init fingerprints; it publishes no code and its own conclusion says the evidence supports, without proving, the inference.
  • 22% → 0.9% — state-framed answers on Nous Research's 803-prompt Propaganda_Eval after its Apache-2.0 fine-tune of Qwen3.8-27B, with fact score rising from 2.32 to 2.98 out of 5 — vendor-self-reported, but with weights, eval, and recipe public for anyone to rerun.
  • +87% after three flat weeks — llama-4-maverick's input price on the eighth OpenRouter Price Index observation, to $0.35/$1.00, while deepseek-v4-flash-0731 fell to $0.0434/$0.30 and a new deepseek-v4.1-flash-1017 row appeared.

1. NEAR AI's breach showed what confidential computing does not cover

The week's lead story arrived in two posts on the same day. First the pause: NEAR AI said on November 9 that it identified unauthorized access to internal systems and paused NEAR AI Cloud inference as a precaution, with its status page showing a major outage — Chat API at 92.59% uptime for the week. Then the report: an attacker used a remote-code-execution vulnerability in SGLang to gain code execution on one inference server and reach an internal system behind deprecated Private Chat. About 100 user accounts (emails, names, and for some accounts plan and balance data) and 18 agents on a legacy non-TEE deployment were affected; two customers' API keys surfaced in logs and were revoked. The report states the vulnerability was reported to the SGLang project and a fix deployed across all inference servers, with no CVE identifier or affected-version range published as of November 14. Why it matters: the checklist every builder uses to buy "private inference" assumed the TEE was the top of the stack. NEAR AI's own report says the enclave protects the data path from the operator and the host — not from the software inside it. The trust question moves from "is there a TEE" to "what code runs inside it." Full story, the outage PULSE.

2. xAI made watermarking the Grok API's default before the EU made it law

The first confirmed default-on watermark policy at a major frontier API landed on November 7, twenty-five days before it was required. Text output from seven Grok model variants — grok-4.7, grok-4.6, grok-4.5, grok-4.3, and three grok-4.20 variants — now carries an invisible statistical watermark on every xAI API endpoint, readable per request through the x-watermark response header. The opt-out exists but is structured like the permission-slip pattern DeAI has tracked all season: a team-admin setting, per model, applying to every key on the team, effective within five minutes, audit-logged, and gated on the admin confirming a legal exception covers the use. xAI's FAQ asserts no quality impact while stating the watermark "slightly shapes which tokens the model picks," and detection is statistical, run by xAI, with access granted by application form only. Why it matters: the EU AI Act's Article 50(2) synthetic-content marking obligation becomes applicable December 2 for systems already on the market — and the first mover chose default-on with an attested opt-out, which is a very different default than the open-weights world, where marking is a choice the weights owner never has to make. Full story.

3. Jensen Huang resigned, and the open-weights hardware layer lost its architect

The week's biggest headline was a governance story with infrastructure consequences. Nvidia's board announced on November 12 that Huang resigned as President, CEO, and board member effective immediately, ending a 33-year run; CFO Colette Kress took the seat on an acting basis while the board searches internally and externally, and Harvey C. Jones became the company's first-ever board chair. The announcement states no reason; it followed roughly a week of press reporting on an extramarital affair and an Nvidia investigation into an alleged Chinese intelligence connection. Nvidia says operations and the product roadmap continue without interruption. For this beat, the stakes are specific: Huang ran the company whose export controls shape which labs can train what, whose $12.93 billion agreement to acquire Hugging Face — announced September 3 with commitments to keep the platform open — has not been rescinded, and which has released more than 500 open models on Hugging Face. Why it matters: the Hugging Face deal's "keep the platform open" commitments were personal-policy under Huang, and open weights' hardware layer now reports to an interim CEO with no permanent successor named. Full story.

4. The checkpoint-forensics report: 3 million views, zero reproductions

The loudest open-weights argument of the week was a report, not a fact. Will DePue published "checkpoint forensics" research on November 12 claiming that six Chinese open-weights models — DeepSeek-V4, Kimi K3, Kimi K3.1-Pro, MiniMax-M3, and Ling-3.1 variants from DeepSeek, Moonshot AI, MiniMax, and Ant Group — carry PyTorch initialization seeds and streaming-multiprocessor thread layouts matching Nvidia Blackwell-class GPUs, implying such GPUs trained the weights despite US export restrictions. The announcement post drew more than 26,000 likes and over 3 million views within a day. The report publishes no code, describes one analysis pass not yet re-run, and its own conclusion states the fingerprint work "supports, without proving, the inference that Blackwell also performed the training." No independent party has reproduced the seed-search results. Why it matters: this is the proof problem in miniature — an export-control-relevant claim traveling at X velocity while the reproducibility artifacts sit unpublished, and the correct posture for a builder is exactly the tracker's: claimed until a rerun exists. Full story.

5. Nous Research's Arete-27B is the counter-argument you can rerun

The same week's constructive release made the opposite choice about evidence. Nous Research shipped Arete-27B on Hugging Face under Apache 2.0: a supervised fine-tune of Alibaba's Qwen3.8-27B, post-trained on about 52.8 million tokens with its Inkling model as teacher, that on the lab's own 803-prompt Propaganda_Eval cut answers containing Chinese state framing from 22% to 0.9% while the fact score rose from 2.32 to 2.98 out of 5. The caveats are real — the figures come from Nous's own eval, graded by a vendor-chosen judge (Grok 4.7), with non-identical decoding settings across compared models, and the reported capability retention (65.60% versus 65.77% on Hermes Bench) is self-reported. But the weights, the eval dataset, and the full training recipe are public, and 38 community quantizations appeared within the week. Why it matters: value-aligned fine-tunes of open foundations are now a category a small lab can ship, and the difference between Arete-27B and the forensics report above is precisely the difference between "check my work" and "trust me" — the eval is public, so the claim is falsifiable. Full story.

6. OpenAI pooled ChatGPT and Codex into one token allowance

The quiet policy change with the loudest billing implications. Starting in December, OpenAI measures usage on Plus, Pro, and Business plans in tokens and draws one shared allowance down across both ChatGPT and Codex, per its help center — so coding-agent traffic now consumes the same pooled allowance as chat, and the unit of measurement changes from "a month of chatting" to tokens. The same month removes six model snapshots (gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3, and o3-pro, plus fine-tuned variants) from the API on December 11, with OpenAI recommending the gpt-5.6 family as replacements. Why it matters: anyone who sized a workload or an agent's budget on flat-rate plan assumptions should recheck the math before December — and for the private-inference and open-weights crowd watching vendor metering grow more granular, this is one more data point that the token, not the seat, is the unit of account. Full story.

7. The price index broke its own pattern again: Maverick up 87% on nothing

The eighth OpenRouter Price Index observation landed Monday with the second consecutive week of mixed-direction movement: eight of seventeen tracked rows moved with no shared pattern. The headline was llama-4-maverick jumping 87% on input to $0.35/$1.00 after three flat weeks — the week's only mover against the "cheap drift" narrative. On the other side, deepseek-v4-flash-0731 fell to $0.0434/$0.30, the cheapest input leg the row has carried and its fourth distinct shape in four observations; deepseek-v4-pro gave back its prior 67% spike; and a new deepseek-v4.1-flash-1017 row appeared at $0.006/$0.50 with no first-party DeepSeek pricing change verified for either row. Kimi K3's output leg rose 6.7% to $16.00, continuing its oscillation between $15.00 and $20.00 since mid-October. Why it matters: aggregator rows are not list prices, and both directions of movement in one week is the strongest signal yet that these are per-host experiments rather than a market trend. Full story.

8. The serving layer got its reference cards

Two evergreen additions landed for builders choosing where to run. DeAI priced Together AI's full serverless card: 25 chat models as of November 10, from GLM-5.3 Flash and Qwen3.8 Flash at $0.15/$0.50 and $0.15/$0.47 to Kimi K3 at $3.00/$15.00, with cached input at 90–98% off on the Flash tiers, Batch API at 50% off serverless rates — and a privacy posture worth reading twice: no training on prompts without opt-in, but prompt storage on by default, with self-serve zero data retention that disables passthrough models when enabled. DeAI also shipped the complete host guide for MiniMax M3.1 Flash: a 428B-parameter MoE (~23B active) with a 1M-token context at $0.15/$0.60 across 13+ hosts, self-hostable from a 252 GB checkpoint on 4 Blackwell GPUs (full 1M context) or an 8×H100/H200 node capped at 262,144 tokens — under a MiniMax Community License whose commercial terms require "Built with MiniMax M3.1-Flash" branding and prior written authorization above $20 million in yearly revenue. Why it matters: between the license threshold and the prompt-storage default, both cards document the same lesson as the K3.1 Pro clause last week — the useful terms are in the license file and the org settings, not the launch post. Together AI pricing, the M3.1 Flash host guide.

9. The verification funnel logged its first positive evidence — and still zero verified

The seventh tracker sweep cycle, covering the 2026-11-02 window, appended 19 claimed entries to 13 of 17 tracked providers and grew the history ledger to 104 entries. For the first time in seven cycles, two checks produced positive evidence: pricing-page fetches for Chutes and Venice, corroborated with SHA-256 hashes of the fetched pages. Neither moved a grade: a provider's own published pricing page is a single-source surface, so both stay at claim level pending the human gate, and the export generated November 5 still reads 0 verified across 17 providers. Of 37 logged checks, 17 attempted a real observation — three Chutes endpoint probes (answered HTTP 429, recorded as not probeable), four attestation checks (no validatable attestation published), and eight OpenRouter lookups with no cross-check surface. Why it matters: in a week when a codeless forensics report reached 3 million views, the contrast is the point — the slowest evidence pipeline on the beat is the only one that tells you what "verified" would even mean. Full story.

Also on the tape

Four items that did not make the top nine but belong in your feed. loglog.ai's Yvligo — a DeepSeek V4.1 Flash LoRA that runs X ad campaigns as a tool-using agent — drew the week's sharpest specialist-vs-frontier argument, with its own AdVal benchmark showing it ahead of Grok 4.7 and Claude Opus 5.5 on 45 held-out briefs; every number is a vendor self-eval, and the traces are public enough to make an independent rerun possible. The week-ahead sweep found no confirmed model releases or EU AI Act deadlines in the November 9–15 window — the one on-beat watch item, a claimed Turing distributed compute network mainnet beta, is a project self-announcement DeAI held at claim status. And DeAI's daily briefs from November 8 through November 14 carry the day-by-day record.


The Friday hub, This Week in DeAI: The Proof Problem, ties the week's breach, watermarking, and forensics stories into the full arc with complete sourcing. Last week's edition: The 8 DeAI stories that mattered, week 45.

Questions

Did the NEAR AI breach defeat its confidential computing?
Not the hardware isolation itself — but the software inside it. Per NEAR AI's November 9 incident report, attackers exploited a remote-code-execution vulnerability in SGLang, the open-source inference engine, and ran code inside one inference server, observing requests in processing including confidential-mode ones. The report states confidential computing protects data from the operator and the host, not against a vulnerability in the software running inside the enclave.
Does the xAI Grok API watermark responses by default?
Yes. Since November 7, 2026, text output from seven Grok model variants carries an invisible statistical watermark on every xAI API endpoint, with a per-model team-admin opt-out that is audit-logged. xAI frames the change as EU AI Act compliance ahead of the Article 50(2) marking deadline on December 2, 2026.
Why did Jensen Huang resign from Nvidia?
Nvidia's board announced on November 12, 2026 that Huang resigned as President, CEO, and board member effective immediately, ending a 33-year run. The announcement does not state a reason; it followed roughly a week of press reporting on an extramarital affair and an Nvidia investigation into an alleged Chinese intelligence connection. CFO Colette Kress is interim CEO.
Have the Blackwell checkpoint-forensics claims been verified?
No. The report claims six Chinese open-weights models, including DeepSeek-V4 and Kimi K3, carry initialization fingerprints matching Blackwell-class GPUs. It publishes no code, describes one analysis pass not yet re-run, and no independent party has reproduced the finding. The report itself says the evidence supports, without proving, the inference that Blackwell performed the training.

Sources

  1. NEAR AI Security Incident Report: SGLang Vulnerability — NEAR AI
  2. NEAR AI Status — major outage in progress, Chat API 92.59% uptime (Nov 3–9) — NEAR AI (status page)
  3. Output Watermarking (xAI developer documentation) — xAI
  4. EU AI Act Implementation Timeline — EU Artificial Intelligence Act portal (Future of Life Institute)
  5. NVIDIA Announces CEO Transition — NVIDIA Newsroom
  6. Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs — Unemployment Labs
  7. NousResearch/Arete-27B — model card, weights, and eval results — Hugging Face (model card)
  8. Understanding ChatGPT and Codex usage limits — OpenAI Help Center
  9. OpenRouter API — list models — OpenRouter
  10. loglogai/yvligo — model card, adapter weights, and eval table — Hugging Face (model card)

About DeAI

DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.

Powered by Morpheus and StrandCMS

Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more about the Morpheus Inference API →

Sponsor disclosure — not editorial

Powered by Morpheus and StrandCMS. Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more →