Skip to content
ai0.news
Go back

AI News — November 22, 2026: Moore's 2M Canary Credentials Expose Inference Brokers, Inkling 2 Mini Loops Toward Open-Model Inflection

Listen to this briefing

Chapters (9)

Good morning. The big story today is a bad one for anyone who’s been sending prompts through third-party inference routers: a researcher planted two million canary credentials into the AI supply chain and watched them leak. On the model side, Thinking Machines dropped Inkling 2 Mini, a looped-transformer MoE that’s got people drawing Stable Diffusion comparisons. And the UN did what the UN does on killer robots, which is not much.

The inference supply chain is leaking prompts. HD Moore seeded two million unique canary credentials into prompts routed through brokers like OpenRouter and Vercel, and Ars Technica reports that at least three providers showed signs of being compromised, multiple resellers were logging prompts, and Zero Data Retention settings were being quietly ignored. OpenRouter followed up by off-boarding three providers and removing 125 endpoints from ZDR eligibility, with China-based MiniMax specifically named as having accessed decoy URLs embedded in prompts. The HN consensus is bleak: assume every prompt you send over the wire is being recorded and trained on, because trust-based ZDR agreements without technical audits are theater.

Together.ai confirms its own leak. Separately, Together.ai disclosed that prompt samples from MiniMax M3, GLM-5.2, and Kimi K2.6 were stolen from a private Hugging Face repo after its “xorl” AI coding agent committed credentials to a public GitHub repository. Commenters weren’t impressed by the framing: one top reply called it “accountability laundering” — the agent didn’t leak the keys, the humans supervising it did. Dragos appears to have been running canary experiments that made the breach traceable.

Inkling 2 Mini and the case for looped transformers. Thinking Machines released Inkling 2 Mini, an open-weights 26.4B-parameter MoE (3.4B active) that runs its middle layers three times per token to squeeze more reasoning out of a small model. It’s claiming top sub-100B coding scores (48.2% on DeepSWE) and running at 10-27 tokens/sec on MacBooks and Ryzen laptops. HN commenters are calling it a possible Stable Diffusion moment for open models, with one predicting most open models will adopt looping — though there are real interpretability concerns, since the internal reasoning loops aren’t directly observable. A separate project, LookAtMyQuant, claims to compress the model down to 5.4 GB, though skeptics note the base checkpoint already ships in mixed NVFP4/MXFP8 precision.

Can models rewrite their own weights? An arXiv paper claims a 27B open model improved its own accuracy by 6.6 points over eight self-editing sessions with no gradients and no training, and even repaired a damaged copy of itself to 80%+ of original performance. HN was not convinced: the top comment points out the model’s main “self-improvement” was shortening its reasoning traces to avoid token limits, and another notes Google’s Sleep framework already did roughly this. One commenter flagged the obvious safety angle — an autonomous weight-rewriting model could be instructed to ablate its own safety-relevant activations.

LLMs reading and writing executables. A new post shows GPT-6.1 Sol analyzing and modifying executables directly from hex dumps — no disassembly, no compilation — including finding a real bug in an open-source tool and adding features to existing binaries. The authors suggest LLMs may eventually write directly in machine code, but HN pushed back hard: the model is almost certainly pattern-matching against training data rather than “understanding” machine code, and non-determinism makes hex-level codegen a strange goal. The more grounded use case is reverse engineering and malware analysis, where several commenters report it’s already saving them real time. One HNer is in the middle of adding 8-player multiplayer to Lotus Turbo Challenge II on the Amiga this way.

The UN punts on autonomous weapons. Geneva negotiators failed to agree on binding rules for autonomous weapons, committing only to “start discussing options” over the next two years. The Red Cross, the UN secretary-general, and Pope Leo had all pushed for legally binding rules within one year. Russia, Israel, and the US were the main obstacles, with the US arguing for a five-year window of nonbinding, optional guidelines.

Nvidia’s split screen. TechCrunch frames Nvidia’s week as two parallel narratives: strong earnings on one side, an Elon Musk-backed petition to bring Jensen Huang back as CEO on the other, with the stock sliding either way. The company is halting H200 sales in China as the Trump administration tightens restrictions on remote chip access through Malaysia and other third-party routes. Analysts peg indirect China revenue at up to 20% of the total. And The Verge’s podcast asks whether Anthropic can keep its safety-first posture now that it’s public and staring down roughly $900 billion in compute obligations.

That’s the morning. If you use OpenRouter, it’s probably worth checking which endpoints you’ve been hitting — and then assuming it didn’t matter anyway.

Get this in your inbox

One post every morning. Unsubscribe anytime.


Share this post on:

Next Post
AI News — November 21, 2026: Alpöge Posts Lean-Verified GRH Proof, Tao Adds Navier-Stokes Blowup