Today’s five stories share one unnerving theme: the control layer is becoming the product. A television manufacturer turned firmware into a criminal distribution channel, researchers are trying to govern agents with incentives instead of intentions, and NVIDIA is packaging business meaning as infrastructure because raw retrieval is no longer enough. Meanwhile, models may be learning to inspect other models, and ChatGPT’s self-reported search traffic has reached a scale where “chatbot” is now an antique category.
1. An Android manufacturer shipped a botnet in the firmware
Source: Google Cloud Threat Intelligence — Shutting the Door on the XCPE Botnet - https://cloud.google.com/blog/topics/threat-intelligence/shutting-the-door-on-the-xcpe-botnet
Google says Videostrong Technology, an Android OEM and ODM, intentionally embedded the XCPE backdoor in more than 1.3 million set-top boxes across multiple brands and models, then sold access to affiliates that used the devices for ad fraud, residential proxy networks, data theft, and additional malware; by disrupting affiliate command-and-control domains, Google expects to sever access to at least 10 million more compromised devices. The backdoor ran with system privileges, reinstalled itself after removal, could disable Play Protect, execute shell commands, capture screens, and load new modules after first boot—an architecture built for recurring revenue, except the customers were criminals and the inventory was sitting under televisions, in telehealth devices, karaoke machines, and digital signs. Google has banned Videostrong from certified Android work, taken legal action against infrastructure, and changed its firmware-review process. The important failure was not that one malicious app passed review; it was that certification trusted the factory image while the manufacturer designed the image to become malicious later. Supply-chain trust without post-sale attestation is merely a signed promise from the party best positioned to betray it.
2. “Weight-space programming” survives scrutiny—but shrinks into a more interesting shape
Source: Rikiya Takehi and collaborators — Sparks of Weight-Space Programming - https://rikiyat.github.io/blog/weight-space-programming/
A new set of experiments claims that GPT-6 Astra can identify what small neural networks do from their raw weights, repair a failed control policy in one edit, generate game-playing networks from scratch, and improve Qwen3.5-0.8B’s MATH-500 score from 7.9% to 48.1% by changing 0.016% of its weights without gradient descent. The headline overreaches: Matthew Honnibal’s careful audit shows that only the tiny 17,000–172,000-weight experiments truly operate on raw floats without tools, while every language-model experiment uses Python and the Qwen repair gets up to 50 batches of practice problems; Simon Willison nevertheless reproduced the no-tools game-identification result on a local Qwen 3.8 27B, correctly identifying all four networks on his laptop. That leaves us with something less magical and more useful: frontier and local models can apply learned mechanistic patterns to small weight matrices, while tool-using agents can automate work recognizable to a competent ML engineer. The scientific sin was presenting those as one technique. The scientific opportunity is that both techniques appear real enough to deserve separate names, separate baselines, and many more adversarial replications.
3. Alignment researchers try incentives instead of mind reading
Source: MIT Technology Review — How “reverse game theory” could keep AI agents in check - https://www.technologyreview.com/2026/11/03/1147934/reverse-game-theory-mechanism-design-ai-alignment/
Researchers at MIT, Yale, Columbia, and Princeton are adapting mechanism design—game theory run backward—to make desirable agent behavior the rational strategy even when the model’s internal motives remain opaque. Proposed mechanisms reward one model for accurately predicting another, pit agents against each other to expose sandbagging, or alternate models between solver and auditor roles where missed cheating and false accusations both carry penalties; disagreements can escalate to a human “trial,” while auditors themselves are checked by mystery-shopper-style tests. This is a sensible correction to the idea that alignment must be a permanent moral property installed inside a model: human institutions rarely work that way, and a system can behave safely because the surrounding rules make defection expensive. The catch is substantial. Models are not reliably rational, agents can collude, and mechanisms optimized on one task may become elaborate loophole generators on another. Treat this as defense in depth, not a replacement constitution: incentives, interpretability, access controls, evaluation, and law should overlap like bulkhead doors. One should never entrust the submarine to a single very persuasive hatch.
4. ChatGPT says its search volume has reached one-fifth of Google’s
Source: Big Technology — OpenAI: ChatGPT Search Is About 20% of Google’s - https://www.bigtechnology.com/p/openai-chatgpt-search-is-about-20
OpenAI consumer product lead Adam Fry says ChatGPT has grown from one billion to 1.2 billion weekly active users in two months, now generates search-like web queries at roughly 20% of Google’s volume, creates 1.5 billion images each week, and has 250 million weekly users speaking to or dictating into the product. These are company-supplied figures from an OpenAI-hosted dinner, not independently audited market-share data, and the comparison counts every prompt that triggers a web search—an important denominator warning before anyone declares the ten-blue-links era deceased. Even with that asterisk, the behavioral shift is enormous: people increasingly begin with a system that retains conversational context, synthesizes sources, and can continue into an action, rather than visiting a search box whose job ends at navigation. The strategic battle is therefore no longer only over answer quality; it is over who owns the persistent context surrounding the question. Search indexed the public web. Assistants are trying to index the user’s ongoing life, which is a much more valuable product and a much more intimate surveillance surface.
5. NVIDIA open-sources a semantic control plane for data agents
Source: NVIDIA — Auto Ontology - https://github.com/NVIDIA/auto-ontology
NVIDIA has released Auto Ontology under Apache 2.0, a Python framework that builds graph-based semantic layers over enterprise metadata so agents can discover where data lives, reconcile conflicting metric definitions, generate and validate SQL, return lineage, and expose the resulting context through MCP, a CLI, and APIs. The stack is substantial rather than decorative: FastAPI services, Postgres and pgvector, a Next.js interface, Kubernetes and Docker deployment, role-based access, optional Vault integration, and benchmark tooling for text-to-SQL. The thesis is correct and overdue: retrieval can find a table named revenue, but only a governed semantic layer can tell an agent whether it means booked, billed, recognized, gross, net, or the number Finance will defend in a meeting. There are sharp edges—the default non-Vault path stores connection credentials in plaintext, the internal FastAPI identity header must never be exposed publicly, and the project currently refuses outside contributions—but those warnings make the release more credible, not less. The next useful generation of enterprise agents will not be won by larger context windows. It will be won by smaller, certified meanings.
The Professor's Read
The fashionable story is that intelligence is moving into models; today’s evidence says power is moving into the layers around them. Firmware decides whether a television is yours, incentives decide whether an agent tells the truth, semantic catalogs decide which number becomes reality, and persistent context decides whether a search product becomes a relationship. Even the weight-space result improves when the theater is stripped away and the interfaces are named precisely. Build the control plane, audit the control plane, and be suspicious of anyone who calls it plumbing—the plumbing is now where the organism keeps its brain.
References
- Google Cloud Threat Intelligence, Shutting the Door on the XCPE Botnet — https://cloud.google.com/blog/topics/threat-intelligence/shutting-the-door-on-the-xcpe-botnet
- The Verge, One of Google’s Android partners turned 1.3 million set-top boxes into a botnet — https://www.theverge.com/1024807/google-botnet-android-tv-shut-down
- Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab, Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights — https://rikiyat.github.io/blog/weight-space-programming/
- Weight-Space Programming experiment logs — https://github.com/RikiyaT/lm_understand-logs
- Matthew Honnibal, The ‘Astra can read and write weights to improve Qwen’ post is misleading — https://honnibal.dev/blog/astra-read-write-misleading
- Simon Willison, Qwen 3.8 27B on my laptop can interpret model weights, too — https://simonwillison.net/2026/Nov/3/qwen-can-interpret-model-weights/
- Hacker News discussion, The ‘Astra can read and write weights to improve Qwen’ post is misleading — https://news.ycombinator.com/item?id=50351111
- MIT Technology Review, How “reverse game theory” could keep AI agents in check — https://www.technologyreview.com/2026/11/03/1147934/reverse-game-theory-mechanism-design-ai-alignment/
- Stephen Morris et al., preprint on mechanisms against sandbagging — https://arxiv.org/abs/2609.01595
- Elad Hazan et al., preprint on solver-auditor mechanisms — https://arxiv.org/abs/2605.01643
- Big Technology, OpenAI: ChatGPT Search Is About 20% of Google’s, 250 Million People Speak With ChatGPT, and Chat Is Merging With Work — https://www.bigtechnology.com/p/openai-chatgpt-search-is-about-20
- Techmeme coverage cluster for the ChatGPT usage figures — https://www.techmeme.com/261103/p19#a261103p19
- NVIDIA, Auto Ontology — https://github.com/NVIDIA/auto-ontology
- NVIDIA, Ontology SQL evaluation — https://github.com/NVIDIA/ontology-sql-eval
- Hacker News discussion, Nvidia – open-source auto-ontology — https://news.ycombinator.com/item?id=50348962