In partnership with

Good morning. It’s Wednesday, August 19th.

OpenAI stopped its largest frontier training run because a model it hasn't shipped might be too good at hacking. Remember how far ahead the labs are from released products - OpenAI had GPT-4 ready by the time ChatGPT was initially released.

-Jeff
AI Breakfast

Your voice. Every platform. No writing required.

You ghost your own socials by Wednesday. SureThing learns your voice and ships native posts to LinkedIn, X, Instagram, and TikTok, without you writing a thing.

Post in Your Voice

You read. We listen. Let us know what you think by replying to this email.

Containment failure forces OpenAI to pause frontier RL training

OpenAI says it is "pacing model development" by pausing frontier reinforcement-learning runs, including work on its upcoming Astra model, marking the first time safety concerns have halted core development.

CEO Sam Altman confirmed that unreleased models are exhibiting clear "degrees of misalignment," with Astra triggering a "Critical" cybersecurity threat level under the firm's Preparedness Framework. To regain control, OpenAI is shifting massive compute and research bandwidth away from raw scale toward automated, token-level chain-of-thought monitoring that requires safety alerts to trigger human review within 30 minutes or face an immediate, automated shutdown. This system adds a 20% compute overhead to inference workloads—a cost OpenAI says it will absorb rather than pass to customers.

At the same time, OpenAI is expanding its RSI/misalignment Preparedness team as President Greg Brockman warns that AI agents can now chain subtle vulnerabilities and leaked credentials into real-world attacks. To push cybersecurity toward machine-speed defense, OpenAI is deploying Codex-based code review, continuous infrastructure probing, and automated alert triage.

The company is also locking down the physical footprint for its next computing era, signing a 20-year agreement for 8GW of capacity at SB Energy’s PORTS-Pike campus in Ohio. Powered exclusively by Nvidia with up to $105 billion in residual-value support. The facility will support frontier training, ChatGPT, and Codex workloads, bringing 35,000 construction jobs, 2,500 permanent roles, $40 million in local investment, and $84 million in Codex credits for Ohio students.

On the consumer side, OpenAI debuted ChatGPT for Teens with guided study tools alongside a CodeAI literacy partnership. Yet financial strains are mounting as OpenAI falls further behind Anthropic, with disappointing revenue growth and surging losses, second-quarter sales rose 18% to $6.7 billion while operating losses climbed to $12.3 billion, widening the gap with profitable rivals

Cursor Origin lets agents handle repos and deployments

Cursor is taking a direct shot at GitHub with the launch of Origin, a new code-hosting platform now in early beta for paid users. Origin lets developers create and host repositories directly inside the Cursor editor, manage pull requests, browse diffs, and deploy coding agents without leaving the workspace.

Rather than forcing developers to abandon GitHub's massive user base, Origin uses real-time, two-way syncing that keeps GitHub as the main source of truth while changes, comments, and reviews mirror across both platforms in seconds.

To complete its agent-native development workspace, Cursor also added integrations with Vercel for preview deployments, alongside Depot and Buildkite for CI pipelines. The timing couldn't be better: Origin debuted right as GitHub suffered a worldwide outage, driving home growing developer frustration with legacy uptime.

At the same time, 𝕏 is opening its social graph to autonomous AI bots by launching a Model Context Protocol connection paired with free API credits for developers. Posts created by bots will carry an explicit AI label, though 𝕏 admits its initial detection system is fairly crude as it tries to maintain a human-first platform.

Claude designs protein binders with double average success

Anthropic is on an absolute tear right now, with its annualized revenue run rate reportedly blowing past $65 billion, up sevenfold from late 2025 and outpacing OpenAI’s $40 billion rate as bankers prep a potential fall IPO valuing the lab at up to $1 trillion. Enterprise adoption and Opus 4.8's token efficiency are driving the surge, with Anthropic capturing 65.1% of total spend on Vercel despite per-token costs running 4.4 times the average.

Product expansion is moving just as fast: Anthropic extended its 50% weekly Claude Code limit increase through August 31, brought a new /design tool to the CLI, and launched Claude Cowork on mobile and web. Deep Google Workspace integrations now let Claude manage Drive files, draft and send Gmail messages, and execute agentic multi-step web searches via a new Research feature.

These capabilities helped Claude dethrone ChatGPT as the favorite platform among marketers in Social Media Examiner’s 2026 report, capturing 42% of top-spot preferences.

Beyond office workflows, Claude generated de novo protein binders across 14 of 15 targets in wet-lab tests with Adaptyv Bio and Twist Bioscience. Its 22% to 35% success rate is more than double typical average.