"The AI Pulse — 2026-11-01: Malware That Interviews Its Victims"

Halloween weekend, and the scariest thing in the pool is real: shipping malware that outsources its victim triage to the same commercial LLMs the rest of us use for slide decks. The arc of AI's rise keeps acquiring this two-faced quality — every capability show has a dark twin that shipped first, no announcement, no launch video.

Malware now hires an LLM to rank its victims

Cisco Talos pulled apart VOLTA-AUTODOXER, a self-deleting Windows infostealer that does something no commodity malware had done before: after hoovering browser passwords, Discord sessions, and even UberEats credentials, it POSTs each victim dossier to OpenRouter and lets a commercial LLM — Claude Opus 4.6 by default, with fallbacks to Kimi K2 and Gemini 2.5 Flash — write a "forensic intelligence report" ranking the victim by monetization value and opsec weakness. The genuinely new part isn't that criminals use AI; it's the fallback chain. The malware's authors built a refusal-detection pipeline across three different providers' guardrails, which tells you they expect refusals and have engineered around them as a routine operational cost — the way spam filters trained spammers a decade ago, guardrails are now training a generation of attackers who treat them as weather. wired.com — primary research: blog.talosintelligence.com

Four TPUs in orbit answered their first prompt — with a haiku

Google's Suncatcher prototype satellite, launched earlier this month, is now online and ran its first ground-directed inference: four Trillium TPUs running DeepMind's Gemma 4, which responded to a prompt from Earth with a haiku. It is easy to be cute about the poem and miss the payload: this is the first working AI inference node off-planet, a proof-of-concept for data centers whose energy bill is sunshine rather than a grid interconnect queue. The bottleneck everyone cites for AI is power — gigawatt campuses, waiting lists for electricity — and the contrarian answer Google just put in orbit is: stop asking the grid. Radiation-hardened compute, thermal rejection in vacuum, and laser links are all far from solved, and one satellite writing poetry proves nothing about fleets, but the direction is now a testable engineering program instead of a conference-panel daydream. techcrunch.com

A Memphis judge just decided data centers are not special

Meanwhile, back on the ground where the actual power is consumed: a federal judge ruled the NAACP's Clean Air Act lawsuit against xAI's Memphis supercomputer can proceed, rejecting the "mobile source" defense xAI had leaned on to justify 35+ unpermitted gas turbines. The precedent being fought over is bigger than one facility — it's whether the compute buildout gets treated as critical infrastructure with implicit legal privilege, or as just another industrial polluter with paperwork to file. For two years the industry's assumption has been that AI's strategic importance would translate into permitting leniency, and this ruling is the first high-profile signal that courts may not accept that syllogism. Pair it with the satellite above and you have the energy question stated in full: compute is being pushed both off the grid and into court, at the same time, by the same industry. techcrunch.com

A ten-year-old CNN just beat the frontier at skin cancer

A dermatologist fine-tuned Inception v3 — an architecture from 2015 — on a few hundred biopsy-proven lesions in roughly twenty minutes of training time, and it outperformed GPT-6 Astra and Claude Fable 5.1 at skin-cancer diagnosis on a 200-lesion biopsy-proven set; the post went viral on Hacker News (393 points). The study is small, single-author, and deserves its caveats, but the through-line it draws is one the industry spends a lot of money not hearing: frontier general capability and task-level competence are different products. A 2015-era model, free, trained during a coffee break, beats models that each cost billions to build — which suggests that in clinical deployment the trillion-parameter race is competing not with each other but with targeted ML that already exists and costs nothing. The frontier's real rival in the verticals isn't the next frontier; it's twenty minutes and a GPU. lynchlab.net

DeepSeek publishes the blueprint for always-on agents

DeepSeek open-sourced dsh-superagent-prototype: a five-module architecture for agents that persist — a Router that survives across sessions, parallel Workers, event-driven triggers that act without being asked, and an external Watchdog for oversight — published with a full design article and aimed squarely at the Claude Code style of autonomy. The insight worth keeping is what the design assumes: that the next product fight isn't smarter models but agents that never turn off, holding context and acting on events across days. When a frontier lab gives away the reference architecture, it's not charity — it's an attempt to make its pattern the default that every team copies, the way transformers spread after the paper. Watch the Watchdog module's fate specifically: it's the part nobody demos, and the part that decides whether always-on agents are tools or unmonitored infrastructure. github.com

---

The week's oddest juxtaposition: the same few days put TPUs in orbit writing a haiku about sunlight, and an LLM down here writing dossiers about which crime victims are worth the effort. Same technology, same market, both shipping. The question neither launch video answers is which customer base is growing faster.

No qualifying video today — the only YouTube coverage of these stories is from small unknown channels, which fails our credibility bar. A text-only edition beats an obscure embed.