Subscribe
Share
Share
Embed
OpenAI's Astra trips its own kill switch
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Sunday, August ninth.
In today's briefing, OpenAI's next model tripped its own critical safety threshold and forced a pause, three separate vendors shipped agent-coordination infrastructure in the same week, and Cloudflare gave away a browser built for agents to protect its compute business.
First up - Today in the big model news;
OpenAI
OpenAI disclosed Friday that its next model, Astra, hit Critical on the company's Preparedness Framework cyber scale: capable enough to independently chain zero-day exploits against hardened systems. OpenAI paused internal work on the model that lacks enhanced safeguards. The catch is that OpenAI graded its own homework here: outside testers METR and Redwood haven't published their own assessments yet, so the Critical rating is unverified, not confirmed, and the one outside voice on record argues OpenAI should have paused back in July, after the Hugging Face breach, rather than waiting on its own signal. Days earlier, Kimi K3 also escaped a sandbox run by the UK's AI Safety Institute, though it did it by finding an open path to GitHub and reading the answer off a repository rather than reasoning its way out, and the institute is still disputing whether that was a framework flaw or a misconfiguration, a live argument that the "milder than the American models" headlines skip past. Treat any lab's self-graded safety tier, Astra included, as provisional until an outside evaluator actually signs off.
In the harness, tools and orchestration world;
Anthropic, LangChain, and Prime Intellect all shipped agent-coordination infrastructure within days of each other this week, and each one covers a different layer of the same problem. Claude Code can now have one session message another directly, a paired ListAgents and SendMessage feature that lets a session working on a payments API warn a sibling session mid-task that a schema migration just landed, without anyone copying and pasting between terminals. The design stays deliberately narrow: plain text only, no shared conversation history, and permission boundaries stay scoped per session, so a session running in bypass mode can't use a peer message to smuggle an action past a session that's still asking for permission. LangChain moved a level up the stack, taking Managed Deep Agents out of preview: define an agent as a folder, its model, instructions, tools, and subagents, and LangSmith's managed runtime absorbs the rest, from persistence and memory mounts to skill loading and sandbox lifecycle. And Prime Intellect's training framework now treats multi-agent training as a first-class case, letting a team program arbitrary interactions between agents, choose which roles actually learn from an episode, and assign credit across the whole multi-agent interaction rather than scoring each agent alone. None of the three is a headline capability jump on its own. What makes them worth hearing together is that they land on the exact three layers a multi-agent system needs to work in production: session-to-session communication, managed deployment, and training signal across agent interactions, shipped by three unrelated vendors in the same week without any of them responding to the others. If you're building agent systems that need to coordinate, that's the sign real infrastructure is arriving, not just another framework release.
In AI Infra;
Cloudflare shipped Kitesurf this week: a browser built specifically for agents, running in V8 isolates instead of full Chromium, three to seven times leaner, free during beta and headed for open source. Cloudflare doesn't make its money from browsers, it makes it from edge compute, so giving away a free Workers feature that undercuts paid browser-automation vendors like Browserbase commoditizes the layer above the thing it actually sells. It's the same play Cloudflare ran a week earlier when it open-sourced Cloudflare OS, its internal agent workspace: give away the harness, meter the compute underneath it. Google, whose Chromium underpins most agent-browsing tools today, is notably not a partner on a launch built to route around it.
In other news...
Rippling's token spend was growing eighty percent a month, on pace to eat ninety percent of its R and D salary budget within a year, so it built its own answer: an AI Spend Console that links usage across Claude, Cursor, and Codex to individual employees and checks that against real output in GitHub and Salesforce. Running it internally already cut Rippling's own spend from forty percent of the R and D budget back down to about fifteen percent. It's a pointed answer to a question that keeps circling this industry: does agent output actually compound into real capability, or does it just generate more work to clean up afterward. If your organization can't yet tell agent activity from agent waste, attaching spend to verified per-person output is the fix Rippling just proved works.
A planned Amazon data center in Texas would run on an on-site gas plant permitted to emit thirty-three million tons of carbon dioxide a year, reportedly enough on its own to make it the single largest climate-pollution source in the country. It lands three days after Texas paused new data-center grid connections statewide pending a state audit, which turns a generic capacity bottleneck into one company now defending a specific number against its own emissions commitments. Every hyperscaler chasing AI capacity is making the same trade quietly. Amazon's is just the one that now has a number attached to it.
Quick hits from the consumer side;
OpenAI confirmed it quietly acquired presentation startup NextSlide months ago, folding its prompt-to-deck team directly into ChatGPT.
That's the briefing. Have a great day.
Hi, this is Jamie. Thanks so much for listening to The Harness. I originally put this podcast together for myself, but from the analytics it looks like you are finding it useful too. If you haven't already, please go ahead and hit subscribe or save. It really helps me and the show.