The day's tech, sifted: Sep 11, 2026

What matters today: OpenAI opened up the Agents API, a managed service on the same Codex harness that powers its own coding agent, letting developers spin up long-running, tool-using cloud agents with a single call; Vercel shipped a day-one integration the same morning. The industry's build-out kept accelerating even as its loudest voice questioned the pace: Sam Altman told OpenAI staff the company is weighing a slowdown in frontier development and hopes rivals follow, while OpenAI separately asked members of Congress whether coordinating an industry-wide slowdown would violate antitrust law, a concrete first move on yesterday's safety-governance warnings from Anthropic's Jacob Coxon and Paul Christiano. Oracle's own numbers argue the money keeps moving regardless: cloud infrastructure revenue jumped 121% YoY to $7.4B, beating estimates and sending shares up more than 6% after hours.

AI / LLMs

• OpenAI opened its Agents API, a managed service running on the same Codex harness that powers its own coding agent, handling orchestration, long-running sessions and tool use behind a single API call; Vercel shipped a day-one integration pairing OpenAI's agent loop and session state with its own Sandboxes for isolated code execution, a persistent workspace and a scale-to-zero architecture.
• Sam Altman told OpenAI staff the company is considering slowing its frontier development and hopes rivals do the same, while OpenAI separately asked members of Congress whether industry-wide coordination on a slowdown would run afoul of antitrust law: a concrete first move on yesterday's safety-governance warnings from Anthropic's Jacob Coxon and Paul Christiano, now running into the legal machinery that makes labs coordinating on anything risky.
• OpenAI's Navier-Stokes release included a Lean 4 formal proof, the first new detail since yesterday's dispute over whether OpenAI raced NYU's Tristan Buckmaster and Anthropic researcher Levent Alpöge's unpublished work to publication; a machine-checked proof raises the bar for anyone still contesting the result, though it doesn't settle who saw what first.
• Cursor's new Projects feature replaces one-off chats with a persistent coordinator agent that delegates to subagents, syncs context across a codebase and runs recurring work, pushing coding agents further from single sessions toward standing infrastructure.
• Sakana AI split its Fugu model into Max and Ultra v2, both orchestrating pools of smaller open models to match frontier LLMs on capability while cutting cost 60%, a bet that coordination beats scale for cost-sensitive deployments.

Devtools & Infra

• Ory Agent Security gives every AI coding agent its own verified identity, so audit logs can finally show which agent acted, under whose delegation and why, a governance layer arriving just as agents like Cursor's coordinator start directing thousands of subagents unsupervised.
• Microsoft plans to more than triple its data center capacity from 12GW today to 38GW+ by 2032, with about a third earmarked for AI-specific chips, after shortages forced it to turn away AI and cloud business this year.
• The Pentagon is in talks to lend roughly $5B to neocloud Fluidstack, an AI infrastructure company advised on the loan by Palmer Luckey's Erebor Bank, another sign of defense financing and the AI buildout blurring together.

Security & Privacy

• Forgejo disclosed a critical remote-code-execution vulnerability affecting versions <=16.0.3; self-hosted instances should patch to 16.0.4 immediately.
• Anthropic published its September 2026 threat intelligence report on detecting and countering misuse of AI, the latest in its recurring series of writeups on how its own models are being abused in the wild.
• LinkedIn beat two "BrowserGate" class actions over scanning users' Chrome extensions, after a federal judge ruled neither plaintiff alleged they actually had an extension that leaked private data to LinkedIn, though he left them room to refile.
• Android can now securely migrate saved logins between password managers entirely on-device, replacing the old workaround of exporting to an unencrypted CSV file.

Startups & Industry

• Oracle's Q1 revenue rose 30% YoY to $19.35B, beating estimates, with cloud infrastructure revenue up 121% to $7.4B; shares jumped more than 6% after hours.
• Adobe's Q3 revenue rose 13% YoY to $6.76B, beating estimates, with AI-first annual recurring revenue up 150% YoY and freemium users pushing total MAUs past 1B company-wide; Q4 guidance came in slightly below estimates.
• California Gov. Gavin Newsom signed bills restricting "addictive" social media features for under-16s and limiting minors' chatbot interactions; EFF immediately pushed back, calling the social-media bill a functional ban for under-16s that cuts vulnerable teens off from support communities rather than fixing platform design.
• A federal judge found TikTok liable for lying about its safety standards and exposing minors to inappropriate content, in a Texas AG lawsuit that now proceeds to trial to determine penalties, the same day California moved to restrict minors' social media use by statute instead of by lawsuit.
• Jimmy Kimmel's interview with Senate candidate James Talarico will air only on YouTube, not on ABC, after threats from the Trump FCC against Disney-owned local station licenses, the latest retreat in a monthslong pattern of pressure on Kimmel's show.

Research

• Nvidia's Nemotron 3 Ultra checkpoints scored 30/42 at IMO 2026, reaching the gold-medal threshold using only natural-language proof generation, no formal prover or external tools, via an iterative search-verify-refine pipeline; the team released the checkpoints, training data and code.

Elsewhere

• Apple's John Ternus and Greg Joswiak walked through the design tradeoffs behind the iPhone Duo a day after its $1,999 debut, including minimizing the crease and their read on Android's existing foldables.

Hacker News

Security news leads: a critical RCE in Forgejo (patched in 16.0.4) threatens self-hosted git instances (release notes, 157 points, covered above), and a writeup on "the Deathray" shows a trivial way for an untrusted website to freeze a Mac (post, 116 points, 75 comments). PlanetScale launched Neki, a sharded Postgres offering (site, 101 points), and an essay arguing "creativity is the new moat" against AI commoditization drew steady engagement (post, 143 points, 81 comments). Today's other AI stories, the OpenAI Agents API, Anthropic's September threat intelligence report, and the Lean 4 proof bundled with OpenAI's Navier-Stokes release, are each covered above.

Beyond tech: Google will buy half the electricity from one of Finland's nuclear power plants (BBC, 137 points, 99 comments), a 2024 paper on Silicon Valley reshaping the military-industrial complex resurfaced with the day's heaviest discussion (paper, 163 points, 324 comments, discussion), and a personal essay on hoarding cables pulled the top vote count of the batch (post, 387 points, 281 comments). A free "Music Theory for the 21st-Century Classroom" textbook rounds out the list (site, 182 points, 81 comments).

Threads

• Racing and slowing shared a headline: OpenAI's Agents API launch and Microsoft's 38GW data center build-out kept the AI capacity race at full speed the same day Altman raised the idea of an industry-wide slowdown and asked Congress if coordinating one would even be legal, a direct answer to yesterday's Coxon and Christiano warnings that the industry itself doesn't yet know how to act on.
• OpenAI's Lean 4 formal proof is the first new fact in yesterday's contested Navier-Stokes priority dispute with NYU's Tristan Buckmaster, though it settles the math, not the credit.
• Child-safety policy split into two tracks the same day: California legislated restrictions on minors' social media and chatbot use while a Texas court found TikTok liable for the harm that kind of law targets, and EFF argued neither approach fixes what's actually wrong with the platforms.
• AI agents gained autonomy and oversight in the same breath: Cursor's coordinator agent now directs thousands of subagents the same day Ory gave every agent its own auditable identity, while the Pentagon's talks to lend Fluidstack $5B show how much of that autonomy is now backed by defense money.