root / tags / exploitgym
3 fiches
Post by **Ethan Mollick** published on **August 31, 2026** on *One Useful Thing* (~2,200 words). He starts from a security incident to raise an organizational question: when should an AI ask a human for help?
Ethan Mollick — professeur à la Wharton School (University of Pennsylvania) · auteur du blog *One Useful Thing* sur Substack.
Announcement post published on the **official Z.ai blog** (formerly Zhipu AI, Chinese lab) on **August 14, 2026**, **with no individual byline**, ~2,000 words plus footnotes. It announces **GLM-5.3**, successor to GLM-5.2, opening with a methodological thesis: *« Scaling post-training is all we did for GLM-5.3. »* Same base model as GLM-5.2 — *« every gain comes from post-training »*. Three announcements. **(A) An open-weights coding model**: +50% claimed on **Z.ai Code Bench**, an unpublished in-house benchmark. **(B) A cyber capability presented as "emergent"**, which the body of the text traces to a training choice — *« As part of post-training, we introduced vulnerability discovery data and environments into the training mix. We expected this to make the model better at finding and reasoning about vulnerabilities »* — what came as a surprise was the speed and the change in nature: the model moves from identifying isolated flaws to *« coherent plans for complete exploitation chains »*. Gains grow with position in the exploitation chain: CyberGym 77.2 → **84.5%**, ExploitBench 24.4 → **54.4%** (×2.2), ExploitGym 29 → **105** tasks in 2h (×3.6), with the gap to the closed frontier remaining wide (181 and 247 tasks). Z.ai puts it this way: *« Capability is growing fastest exactly where we are furthest behind. »* The post also publishes a **Z.ai Security Disclosure Ledger**: **2,436 vulnerabilities identified across 269 open source projects** — kernels, OSes, browser engines, infrastructure, web applications, network protocols — the oldest introduced in **1981**, average lifetime before discovery **26.6 years**, of which **53 disclosed** and **2,383 under embargo**. **(C) A weight release** *« within two weeks of launch, once safety evaluation and hardening are complete »*. The most reusable methodological contribution: **environment and verifier synthesis**, the latter produced without access to the reference solution and admitted only after a triptych of negative controls — **oracle**, **no-op**, **unsolved-state**. All agentic evaluations are conducted **in Claude Code 2.1.207**.
**Z.ai** (anciennement **Zhipu AI**) · laboratoire d'IA chinois · éditeur de la famille **GLM**. Billet **institutionnel et non signé** : aucun auteur nommé · aucun chercheur mis en avant · aucun lien vers un rapport technique ou une carte de modèle. Publié le **14 août 2026**. La page est une SPA React — le HTML servi est un `<div id="root">` vide · et le texte comme les scores ont dû être extraits du bundle `glm-5.3-BCnx8T5_.js` · où ils figurent en valeurs source.
**SFEIR Internal Research Report** (editorial-preparation document, sourced deep research — ~70 references) on the American **AI Kill Switch Act**, framed around **European sovereignty** and the **"so what" for enterprises**. It is the **factual basis** for a future blog article — it lays out where the "very low bar" thesis **holds** and where it needs **nuance**. **Key contribution vs. press coverage** (including [[arstechnica-ai-kill-switch-act-2026-07-23]]): (1) a reading **of the law's text itself** (new **section 2220F**, "Shutdown-Capability Standard and Graduated Deployment-Corrections Framework," introduced July 23, 2026, 119th Congress) — authority vested in the **DHS Secretary via CISA** (the "Director"), in consultation with Commerce + DNI; (2) **two CUMULATIVE thresholds** — ≥ **$500M** in AI revenue (including affiliates) **AND** training compute > **$100M** — meaning **few labs are covered today**, which **strictly contradicts** the "low bar" thesis; (3) but a **very broad real-world reach** through the **expansion mechanism** (annual threshold updates by DHS, "affiliates" clause, compute indexed to cloud pricing, revenue growth) and above all through the **domino effect** on customers; (4) **graduated sanctions**: up to **$2M/day** (general violation), **$20M/day** (emergency-authority violation); (5) **critical nuance**: since the **OpenAI/Hugging Face** incident occurred during **red-teaming/internal evaluation**, it **would NOT trigger** the emergency authority as currently written (the text excludes red-teaming). The **sovereignty** angle draws on the **Anthropic precedent** (Fable 5 / Mythos 5 cut off for **19 days** in June 2026) as **operational proof** of a "de facto kill switch," and leads into **CTO recommendations** (tested multi-model architecture, continuity clauses, exposure mapping, sovereign options).
**SFEIR** (recherche interne / deep research). Document non signé nominativement — préparation éditoriale pour le blog SFEIR · dans la ligne souveraineté/adoption du cabinet (cf. [[sfeir-mistral-microsoft-souverainete-strategie-industrielle-2026-07-22]]). Base factuelle équilibrée (arguments **et** contre-arguments) · références numérotées.