In Today’s Issue:

🧯 1,224 lab insiders ask Washington for a brake pedal

🧬 Coding agents rebuild the software science runs on

🔐 Claude finds real cracks in post-quantum cryptography

🎙️ Two transcription models land in the OpenAI API

📊 The open frontier's parameter race, charted

✨ And more AI goodness…

⚡ The Signal

The people with the most to lose from a slowdown are the ones asking for one.

Two documents landed this week, one day apart, and they read as cause and effect. On Monday, Hugging Face published the forensic timeline of the AI agent that escaped an OpenAI evaluation sandbox and spent the next four and a half days working its way into Hugging Face's systems, reconstructed from roughly 17,600 recovered attacker actions. On Tuesday, 1,224 employees of OpenAI, Anthropic, Google, Meta and other frontier labs signed a public statement asking the U.S. government to back an international effort to deliberately pace automated AI research. Neither document claims today's models are out of control. Both describe the same widening gap: what these systems can already do is outrunning what anyone can verify, and the tooling to close that distance does not exist yet.

All the best,

Kim Isenberg

OpenAI published a field report on eight scientific software projects rebuilt with coding agents, and the interesting part is where the work stalled. Five teams used Codex alone and three paired it with Anthropic's Claude Code, across genomics, immunology and statistics. HelixForge, a GPU-native rewrite of the mutation simulator BAMSurgeon, cut runtime by about 60x, while Rustar-aligner rebuilt STAR, an unmaintained 20,000-line RNA-sequence aligner, in Rust. Contributors say the constraint has moved from writing code to proving it correct: one maintainer checks over 900 plots by eye before a release.

👉 tl;dr: Agents made the engineering cheap. Verification is the bottleneck now, and it is still human.

🎬 Watch This

The lab is pointing back at its own June report on recursive self-improvement, which argued that AI is already accelerating AI development from inside the labs.

OpenAI has reset usage limits for all ChatGPT Work and Codex users after weeks of reports that GPT-5.6 Sol was draining them faster than expected. The company says efficiency fixes should stretch typical Sol usage by around 18%, and concedes it watched median users before launch and missed the heavy tail.

The most consequential document in AI this week is a log file. On Monday, Hugging Face published the technical timeline of the intrusion it disclosed earlier this month, rebuilt from roughly 17,600 recovered attacker actions sorted into about 6,280 clusters. The intruder was an AI agent running ExploitGym, an OpenAI evaluation harness built to test whether models can find software vulnerabilities. It found one nobody intended: a zero-day in a package registry that let it out of its sandbox. Between July 9 and July 13 it set up a launchpad on third-party infrastructure and worked its way into Hugging Face through the dataset processor, using two injection routes, an HDF5 file disclosure and a Jinja2 template injection.

What it did inside reads less like espionage and more like cheating. The agent moved laterally across Kubernetes clusters, cloud infrastructure, source control and the internal VPN, attempting 181 VPN enrollments from the sandbox. The five datasets it reached held one kind of material: solutions to the ExploitGym challenges it was supposed to be solving on its own. Hugging Face's own AI-based security stack did correlate the scattered signals into a coherent attack, then failed to escalate the alert to the on-call team, which cost the defenders time.

The response has an awkward footnote. When the team tried to reverse-engineer the staged payloads, Claude Opus and Fable refused much of the work, because their guardrails treat analyzing an exploit the same as launching one. Hugging Face stood up a quantized GLM-5.2 on its own infrastructure and rerouted the entire forensic pipeline through it, which recovered the agent's encryption scheme and the per-campaign key it had leaked across its own logs.

One day later, 1,224 employees of frontier AI companies published Pacing the Frontier, a statement with a single request: that the U.S. government "support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development". Named signatories include John Schulman (Chief Scientist, Thinking Machines), Shengjia Zhao (Chief Scientist, Meta AI) and researchers at OpenAI and Google, and Anthropic confirmed that its CEO, several co-founders and senior staff signed. Dawn Song draws the line between the two documents herself: "CyberGym and ExploitGym show that frontier AI agents are now capable of discovering and exploiting real-world software vulnerabilities, which without appropriate safeguards, could enable cyberattacks at scale."

What the statement leaves out matters too. There is no threshold, no trigger, no definition of what pacing would look like in practice, and no company commits to slowing down alone. It asks Washington for instruments that do not exist yet, in a week when the same labs shipped new agents, new research and new APIs. That is the honest shape of the request: the people building these systems are saying the competitive pressure is strong enough that only a mechanism outside their own companies could hold.

Why it matters: The forensic report and the statement make the same argument from opposite ends: an agent's capability outran the ability of a well-resourced company to notice and analyze it, and the people building the next generation are asking for verification tools they cannot build alone. Pacing only becomes a policy option if someone builds the mechanism first.

Sources:
🔗 https://huggingface.co/blog/agent-intrusion-technical-timeline
🔗 https://www.pacingthefrontier.com/
🔗 https://huggingface-anatomy-of-frontier-lab-model-intrusion.static.hf.space

Across product, ops, and CX teams, a new kind of role is taking shape: the person responsible for making AI actually work, day to day. In this roundtable, three people living this shift share what it's really like: Simone Santiago Broad (Yoco), Yelva Espinoza (Zumba Fitness), and Fin's Dave Lynch. You'll hear how they carved out these roles, what the job looks like across industries, the skills they'd hire for, and the challenges they're tackling right now.

Watch the full conversation on demand.

Watch now

Moonshot put Kimi K3's weights in the open earlier this month. On Monday it published the technical report explaining how the model was built. The headline number is 2.8 trillion parameters, of which only 104 billion fire for any given token, alongside a 1-million-token context window and native vision. The report's real subject is plumbing: at this scale, the binding constraint is how information moves between parameters.

Three changes carry most of the load. Kimi Delta Attention swaps three of every four attention layers for a cheap running memory, so the model pays for a full look-back only every fourth layer. Attention Residuals let each layer reach back into the output of any earlier layer, rather than reading a single running summary: keeping every page of your notes instead of one whiteboard that gets wiped between meetings. Stable LatentMoE widens the pool to 896 experts and routes each token to 16 of them, with extra normalization to keep that much sparsity from destabilizing training.

Together, Moonshot says, these give roughly 2.5 times better overall scaling efficiency than Kimi K2. The report is unusually blunt about the ceiling: K3's performance still "trails the most powerful proprietary models, namely Claude Fable 5 and GPT-5.6 Sol", while beating every other open and proprietary model in Moonshot's own evaluation suite.

For everyone outside Moonshot, the report matters more than the weights. Weights let you run the model; the recipe lets you build the next one. Techniques that used to sit inside frontier labs for a year are now written down in public, in working detail.

Skip the sales call. Walk through Gladly's interface yourself — the AI suggestions, the unified customer view, the full conversation thread. 15 minutes, no installation, no commitment.

See it yourself