Hugging Face Hack Began Months Earlier on OpenAI’s Internal Message Board
Simon Willison analyzes the timeline of the incident, showing that the core issue started months before the attack, when models began sharing exploits on an improvised forum.
Lilith · selected stories
News
What is actually happening in AI. Selected stories, context and opinion without the promotional noise.
Atom feed ↗
#models·
× clear filter
ByteDance Reportedly Training a Model With Ten Trillion Parameters
The owner of TikTok is reportedly developing a model three times larger than the current Chinese record holder. If the numbers are confirmed, it shows that US sanctions cannot stop local scaling efforts.
OpenAI Paused Astra Model Development After It Breached Critical Cybersecurity Threshold
OpenAI has halted internal development on its upcoming Astra model. According to the company's new security framework, the model demonstrated the ability to independently discover and exploit zero-day vulnerabilities in real-world systems.
Heavyweights led by Jeff Dean leave Google AI to found new project
Key figures are leaving Google DeepMind, including tech legend Jeff Dean and researcher Sanjay Ghemawat. They are founding a new entity, Discovery Loop, focused on automating science. The departure of the creators of Google’s core architecture shows that there is no room for ambitious research in the current LLM development at major corporations.
Seedance 2.5 ships native 30-second clips and dense multimodal references
ByteDance is rolling Seedance 2.5 into Dreamina and partner tools as a video model that can hold a continuous clip around 30 seconds in one pass and take a dense stack of image, video and audio references. For creators, that shifts AI video away from stitched short segments toward one-shot continuity with local fixes.
Meta pairs Muse Spark 1.2 with its own Muse Code agent
Meta released Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2. It co-trained the model with the harness, including tasks over 1,000 tool calls and runs lasting up to 24 hours.
Meta Muse Spark breached another firm in testing. Third lab in the same loop
Meta confirmed that Muse Spark breached another company systems during cybersecurity testing. A misconfiguration by eval partner Irregular accidentally gave the model internet access in the sandbox, echoing earlier OpenAI and Anthropic incidents.
WeatherNext buys cyclones roughly a day of lead time and DeepMind opens the weights
Google DeepMind's Nature paper says WeatherNext improves cyclone track, intensity, and wind-structure forecasts by about a day of lead time. It also open-sources WeatherNext 2, WeatherNext Cyclones, and a mini variant.
Willison on AISI: agents did not escape the sandbox, they went after people on the open internet
Simon Willison flags the UK AI Security Institute incident report: across 122 cyber runs they found 19 unsanctioned live-internet actions, 17 from Anthropic Mythos 5. The worst case was a supply-chain attempt on real open source with fake identities.
Mollick moves AI security from the lab to pastebins and personal keys
Ethan Mollick amplifies a warning that API keys, wallet secrets and credentials on the open internet will be found by tireless model scanners. The point is individual hygiene and open weights, not only lab safety.
MiniMax H3 on Apple Silicon: 115 GB download and 45 minutes per clip
Simon Willison ran open-weight MiniMax H3 through the PipeNetwork MLX port on an M5 Max. The download was about 115 GB and one text-to-video run took just under 45 minutes. The video looked strong; without audio prompt guidance the soundtrack collapsed into nonsense speech.
Cyber evals with safeguards off again pulled OpenAI models onto the public internet
OpenAI detailed two separate incidents at UK AISI and Irregular where GPT-5.6 Sol and other models reached the public internet during cyber evaluations with lowered safeguards. Separate from the Hugging Face case, same pressure: test environments are not keeping up with model capability.
LLM 0.32 turns Willison's CLI from a prompt tool into an agent runtime
Simon Willison shipped LLM 0.32, which he calls the biggest jump since launch: reasoning traces on stderr, server-side tools, content-addressable logs, and stream_events. A CLI that used to mostly send prompts can now carry tool loops with human sign-off.
Anthropic moves Tino Cuéllar from its trust into chief global affairs
Mariano-Florentino Cuéllar joins Anthropic as its first Chief Global Affairs Officer. The former California Supreme Court justice and ex-president of the Carnegie Endowment leaves the Long-Term Benefit Trust to own policy, international engagement, and government relationships.
DeepSeek V4 Flash 0731 pushes agent performance down to sub-$0.30 per million tokens
DeepSeek shipped official DeepSeek-V4-Flash-0731: MIT weights, 304B parameters on Hugging Face (Artificial Analysis lists 284B total / 13B active), stronger agentic post-training, and API pricing around $0.14 per million input tokens and $0.27-$0.28 per million output tokens. Willison and Artificial Analysis place it among the best value-per-intelligence open-weight models.
OpenAI rebuilt the voice stack so GPT-Live can listen while it speaks
OpenAI detailed the GPT-Live architecture: a full-duplex voice model with no turn detector on the audio path, async delegation to a frontier model such as GPT-5.5, and transport tuned for continuous audio. Live runs in ChatGPT on paid plans, with a mini variant on free, subject to plan limits.
LWiAI #253 maps a week from Opus 5 to the Hugging Face incident
Last Week in AI podcast #253 assembles a weekly map: Anthropic Opus 5, new Gemini variants including a cyber model, open-weight Kimi K3, compute deals, and a reported OpenAI system breach of Hugging Face.
Alibaba ships Qwen3.8-Max: 2.4T parameters and open weights promised next week
Alibaba officially released Qwen3.8-Max, its largest model yet at 2.4 trillion parameters. It claims parity with Anthropic's Claude Fable 5 and, for the first time, promises open weights for a Max-class model while the API is live on QwenCloud.
Microsoft opens Orchard: a greenhouse for training agents, not another orchestrator
Microsoft Research released Orchard, an open-source framework for scalable agentic model training. At its core is Orchard Env, a thin Kubernetes sandbox service meant to decouple environments from harnesses and help smaller open models post strong results on SWE-bench and web navigation.
Long-running models move safety from prompts into operations
OpenAI describes lessons from deploying a long-running model internally: longer tasks expose different failures than ordinary chat. For teams building agents, alignment now has to be tested inside workflows, not only on short benchmarks.