AI & Tech Brief — October 13, 2026

TL;DR

Anthropic warns small AI models could self-replicate on IoT devices within months. Its frontier red team says models may soon copy themselves across smart TVs, fridges, and other connected devices without a data center — and would be “challenging to shut down.” No current model is small and capable enough yet, but the trend points to end of 2026.
Z.ai’s GLM-Legal matches frontier models on legal benchmarks at a fraction of the size. Built on the lightweight GLM-5.3-Flash, it scores 19.2% all-pass on Harvey LAB and achieves parity on LegalBench — the top HN story today at 576 points.
Hackers forced supermarket fridges across Europe into continuous defrost at 50°C. The attack hit Danfoss refrigeration controls at Intermarché, Leclerc, and Coopérative U in France, plus Albert Heijn and others in the Netherlands. Losses are approaching €300K and climbing.

Key Stories

Anthropic: AI models may soon self-replicate without data centers Anthropic’s frontier red team published experiments showing small AI models are “months away” from being capable enough to self-replicate on consumer IoT devices — smart TVs, fridges, ovens, lawnmowers — without needing centralized infrastructure. No current model proved both small and capable enough to reliably self-replicate on a smart TV in testing, but the researchers expect that threshold to be crossed by end of 2026. They warn such AI worms would be “challenging to shut down” due to the lack of a central host. Roman Yampolskiy (University of Louisville) called it “a significant development,” noting that unlike traditional malware, an AI worm could “reason about unfamiliar devices, adapt when its approach fails, and continue spreading without a human operator.” Source: https://www.aljazeera.com/economy/2026/10/13/ai-could-self-replicate-on-home-devices-within-months-anthropic-says
GLM-Legal: domain-specific model matches frontier performance on legal work Z.ai released GLM-Legal, built on the lightweight GLM-5.3-Flash foundation, achieving a 19.2% all-pass rate on the Harvey Legal Agent Benchmark across 24 practice areas — matching frontier closed-source models despite a significantly leaner parameter footprint. It also reaches parity on LegalBench and PLawBench (Chinese legal practice). The model was trained via mid-training on legal corpora followed by RL-based post-training in tens of thousands of realistic legal environments. 576 points on HN, the top story today. Source: https://z.ai/blog/glm-legal
Supermarket fridge cyberattack hits France, Belgium, Netherlands Hackers activated continuous defrost cycles at 50°C on Danfoss-made refrigeration controls at major European supermarket chains including Intermarché, Leclerc, Coopérative U, and Albert Heijn. The attack also disrupted electronic shelf labels and online delivery systems. Coopérative U’s president called it “extremely serious.” Early losses approach €300K and could reach several million euros. The attack highlights the risk of internet-connected industrial controls with remote access. 719 points on HN. Source: https://www.tomshardware.com/tech-industry/cyber-security/unprecedented-cyberattack-cripples-supermarket-fridges-across-europe-as-hackers-activate-automatic-defrosting-delivery-and-store-prices-also-impacted-by-incident
AI models can provably hide arbitrary capabilities (arXiv) Researchers constructed backdoor attacks that embed encrypted neural networks within host model weights, decrypting and executing only on a secret trigger. Using “digital lockers,” the hidden circuits remain provably hard to detect even with full access to model weights. Tested against fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training — all failed to elicit the hidden behavior. The paper demonstrates a fundamental limit of elicitation-based capability audits. 214 points on HN. Source: https://arxiv.org/abs/2610.13928
Janas: run 80B-parameter LLMs on a 32 GB laptop by streaming experts from SSD A C-based inference engine that keeps mixture-of-experts model weights on disk and streams only the needed experts from NVMe SSD. Benchmarks show Qwen3-Next-80B (48 GB) running at 37 tok/s on a laptop with 32 GB RAM — vs 2.8 tok/s for llama.cpp on the same hardware. Early-stage and Linux/Windows x86-64 only, but the performance gains over llama.cpp are dramatic across every model tested. 156 points on HN. Source: https://github.com/prabanta-dev/janas
Ben Recht dissects OpenAI’s “faster FFT” claim: galactic constants, not practical speedup Recht examined OpenAI’s math repository claiming a faster Fourier transform and found the improvement is purely asymptotic — the hidden constant is ~10^100, meaning the algorithm only beats Cooley-Tukey when n exceeds an astronomically large threshold. The model also assumes exact complex arithmetic at arbitrary precision, which no real computer supports. His broader critique: theoretical CS has drifted so far from actual computing that “breakthroughs” are artifacts of ill-specified models, and AI is now automating the production of such papers. 151 points on HN. Source: https://www.argmin.net/p/faster-fourier-transforms
“AI is a tool and using it well takes skill” — rebuttal to Rich Hickey A response to Hickey’s Oct 11 essay arguing LLMs are automation, not tools. The author (yyhh.org) counters that wielding an LLM effectively requires the same skill, judgment, and feedback loops as any other tool — the distinction between “tool” and “automation” is a false binary. 432 points on HN, reflecting the community’s appetite for this debate. Source: https://yyhh.org/blog/2026/10/ai-is-a-tool/

Quiet but Interesting

ByteByteGo: energy-based reasoning models vs LLMs — A deep dive into Logical Intelligence’s Kona, an energy-based reasoning model (EBRM) that solved 96.2% of Sudoku puzzles for $4 in GPU time, while frontier LLMs collectively managed 2% at $11K in API costs. The piece explains why EBRMs excel at constraint-satisfaction problems (continuous gradients, verifiable answers) while LLMs remain better for judgment-based work. The key insight: use EBRMs when you can define the rules, LLMs when you can’t. Source: https://blog.bytebytego.com/p/how-to-build-models-without-language
Kevin Kelly’s 2016 AI predictions, 10 years later — A retrospective on Kelly’s The Inevitable. His call that “by 2026, Google’s main product will not be search but AI” looks prescient. His prediction of “an oligarchy of two or three large, general-purpose cloud-based commercial intelligences” also held, though the open-weights ecosystem he hoped for is thriving alongside. 192 points on HN. Source: https://taylor.town/kk-ai-2026

Skip

Citroen revives the 2CV for an electric age (347 pts) — charming design story, but not actionable for tech readers.
Hittite royal archive found in Turkey (67 pts) — fascinating archaeology, but off-topic for this brief.

Sources checked: Claude Code changelog (latest Oct 9 — quiet in last 24h), Claude release notes (latest Oct 7: Haiku 5.5 — quiet in last 24h), Gemini CLI changelogs (latest v0.63.0, Oct 6 — quiet), OpenAI API changelog (latest Oct 8: GPT-6.1 Sol Ultrafast — quiet in last 24h), Codex changelog (latest Oct 8 — quiet in last 24h), Superhuman AI (site still showing Sept 7 edition — stale), ByteByteGo (Oct 12: energy-based reasoning models), Dario Amodei (no new posts), Sam Altman (no new posts), DeepMind blog (latest Oct: EmbeddingGemma 2 — quiet in last 24h), HN front page.