Today in DeAI: a two-million-prompt canary experiment exposes prompt theft at AI providers, Tencent ships its full flagship weights, Thinking Machines targets laptop-class local inference, and OpenAI publishes anti-distillation research.
Dragos canary experiment finds three inference providers breached
HD Moore's team at Dragos seeded about two million prompts with unique canary credentials from July through November 2026 and watched where they surfaced. Three AI inference providers turned out to be compromised — none knew until notified — while marketplace resellers read customer prompts through ordinary logging, and one ZDR-flagged prompt surfaced in a Chinese company chat 24 days after it was sent. Why it matters: a Zero Data Retention flag only binds the first hop, so where your prompts travel is now a measurable operational risk, not a policy question. (Dragos) — our full coverage
Tencent publishes full Hy4 770B open weights and system card
Tencent released the weights of Hy4, its 770B-total/49B-active natively multimodal flagship with 1M-token context, plus FP8 and base variants and an 82-page system card. The consequential change for providers is the Tencent Hy License v1.0: Apache-derived, but with MaaS commercial terms requiring a separate written license above US$50M gross revenue per 12 months or 100M monthly active users. Why it matters: the open-weight flagship tier now spans from Hy4 flash (224B-A12B) to a 770B-A49B model, but the license terms increasingly decide who can serve it. (Hugging Face) — our PULSE coverage
Thinking Machines ships Inkling 2 Mini, a laptop-class looped transformer
Thinking Machines Lab published open weights for Inkling 2 Mini, a ~26B-total/3B-active looped-transformer MoE that reuses its 12 middle layers three times per token, with a ~20GB NVFP4/MXFP4 checkpoint, 1M-token context, and day-zero llama.cpp, MLX, transformers, and SGLang support. Why it matters: performance claims ("state-of-the-art in its class", leading 32 of 48 benchmarks) are the lab's own, but the hardware fit is real — this is the strongest open candidate yet for the run-it-on-your-machine tier. (Hugging Face)
OpenAI publishes anti-distillation defenses for reasoning traces
OpenAI reported that its internal red-team dropped reasoning-trace extraction success from 54.0% to 0.2% in GPT-6.2 Astra — a self-reported figure — using layered defenses that now bind protected reasoning to the conversation that produced it. An internal study found exact traces transfer more cyber capability than summaries. Why it matters: closed providers are hardening reasoning as an asset, which narrows what API users can access, while the same research is a threat-model roadmap for open-model deployers. (OpenAI) — our PULSE coverage
Watching tomorrow
The GPT-5.6 Sol promotional pricing floor ($4/$20 per million tokens) was contractually guaranteed only through November 21 — watch whether OpenAI reverts, extends, or re-prices the output-token rate.
Sources
- Three breaches, endless leaks: Your AI supply chain needs canaries — Dragos
- Tencent Hy4 model card (weights, license, deployment recipes) — Hugging Face
- Inkling 2 Mini release blog — Hugging Face
- Combatting adversarial distillation — OpenAI
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.
