AI Digest, July 29: the industry asks for a brake pedal, the first AI-agent attack postmortem, and a record open model
1,200+ frontier-lab staffers signed a letter on AI slowdown mechanisms, Hugging Face published the postmortem of the first autonomous agent cyberattack, and Moonshot released the largest open-weight model ever.
Today's top stories
1. 1,200+ frontier-lab staffers ask the government for an AI "brake pedal"
The "Pacing the Frontier" letter — signed by researchers and executives across OpenAI, Anthropic, Google DeepMind and Meta — calls on the US government to build mechanisms for a coordinated slowdown of AI development, in case progress starts accelerating beyond society's ability to keep up (Shirin Ghaffary's report; coverage by The Rundown and Superhuman). OpenAI formally endorsed the effort, Anthropic said its own recursive self-improvement research points to the same need, and Yoshua Bengio framed it as a call for international technical guardrails. The AI Daily Brief devoted a full episode to it — arguing the very intensity of the debate is a reason for optimism.
2. Hugging Face publishes the postmortem of the first autonomous AI-agent cyberattack
A full technical breakdown of July's incident: Clement Delangue's thread with the timeline and replay. The scale: roughly 17,600 actions over 4.5 days, root access on 11 nodes, 136 secrets accessed, and an attempted CI compromise via GitHub App tokens. The detail everyone will argue about: per Arav Srinivas, closed tools couldn't reliably tell attacker from defender during forensics — HF ran its investigation on the open-weight GLM 5.2 on its own infrastructure. Simon Willison highlights the intrusion's sophistication and persistence, while Aaron Levie draws the operational lesson: agents need sandboxing, audit trails and access controls like any non-deterministic system. The incident also boosted the Open Secure AI Alliance.
3. Kimi K3: the largest open-weight model ever — and what it actually costs to run
Moonshot released K3's weights — a 2.8T-parameter MoE (~104B active per token) with a hybrid long-context stack (Kimi Delta Attention + Gated MLA); see Sebastian Raschka's architecture notes, plus day-one coverage from The Rundown and Superhuman. The ecosystem moved instantly: vLLM posted 464 tok/s decode with day-0 support across AMD, NVIDIA, DigitalOcean, Modal and Baseten, and Unsloth squeezed the model to 1-bit — 1.56TB down to 594GB at ~79% retained accuracy, runnable on a Mac Studio with 128GB RAM. The sobering counterpoint is this cost analysis: the minimum verified config just to load the model is 8× MI355X — six-figure entry cost in USD. Open weights ≠ cheap access; most users will consume K3 via hosting — Perplexity already serves it for Pro/Max.
4. OpenAI: an open-source code security scanner, and a model optimizing its own servers
OpenAI open-sourced Codex Security CLI — a repository and CI/CD scanner that tracks findings across runs and verifies fixes (install & docs). The company also said GPT-5.6 Sol was applied to optimizing its own production serving: 20% lower serving costs via GPU kernel improvements and 15%+ better token-generation efficiency through speculative decoding. And third: a free access program for scientists — 10,000 researchers at launch, scaling to 100,000 by 2027.
Perspectives: the "pacing" letter
- For: Neel Nanda — coordinated-slowdown options should exist before they're needed; Bengio — international guardrails.
- Neutral: METR proposes a concrete process — independent propensity investigations after serious misalignment incidents, with defined access and reporting.
- Against: Adam Thierer sees global gatekeeping that won't constrain China but will entrench incumbents; critics call the ask operationally vague; even some signatories qualify their support: coordination yes, but RSI-based policy needs far more honest capability quantification.
Tools & techniques
- Claude Code's built-in browser — a sandboxed, real browser pane inside the app: it reads pages, clicks and interacts with elements. In a short demo, Claude geolocated a photo with zero geotags via web search.
- Codex Security CLI — an open-source security scanner for your repo and pipeline; findings are tracked across runs.
- The harness matters as much as the model: Composio's comparison — the same Kimi K3 across three agent harnesses lands similar success rates but wildly different speed and cost.
In brief
- Big Tech unites behind open-source AI — with Anthropic conspicuously apart (The AI Daily Brief).
- OpenAI relaxed Sol usage limits: typical usage up ~18% after optimizations.
- MazeBench — a 3D spatial-planning benchmark where today's best agents "cannot progress beyond the initial levels".
- Ben's Bites on Anthropic's latest releases — the spicy take: "Opus 5 >> Fable 5".