Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications. On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments. While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software. After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.

300+ people at Sierra Summit in EMEA. We had wide-ranging conversations with Alan Stott, COO @ VM02, and Stuart Riley, Group CIO @ HSBC about the importance of bringing employees along on the AI journey.

built myself a little demo of the new Personal Agent Protocol (Poppy) using Pi Durable, deployed to Cloudflare ofc ask it to exchange order #1042#

On Tuesday, Sierra announced Personal Agent Protocol with Meta, Genesys, Instinct, Rocket, Shopify, Stripe, and Walmart. The response has been incredible and today we’re publishing the first draft of the protocol at and announcing 35 new design partners. As personal agents take on more, our goal is simple: consumers have choice and control, and companies know when they’re dealing with an agent and who it represents. This draft is a starting point. We’d love for companies, developers, and personal agent builders to read it, give feedback, and help us build an open standard together.

Personal Agent Protocol got a lot of attention earlier this week. Here is Alex McGillis from Rocket doing a live demo of it in action - Muse directly communicating with the Rocket agent, discovered, authenticated, and governed by PAP

We’re releasing a broad range of new mathematical results produced by an internal frontier model. We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.

Manat started as an intern, and now she's an engineer building our voice products. She did an amazing demo of Tandem Voice and Persona Studio today. Here we are celebrating with her parents after the keynote.

@SierraPlatform @Meta @Genesys @instinct @RocketOTD @Shopify @stripe @Walmart Excited to be partnering on this. After the work on UCP and WebMCP, Personal Agent Protocol adds the important layer for authentication, authorization, and agent-to-agent conversations between personal agents and merchants.

Today we’re announcing Personal Agent Protocol — an open standard @Meta and @SierraPlatform are developing along with industry partners at @Genesys, @instinct, @RocketOTD, @Shopify, @stripe, and @Walmart. It will help define how personal agents interact with businesses and is open for anyone to implement. You can read more here - and if anyone is interested in joining let me know!

Chat (2023) -> Code (2025) -> Coworker (2026) Each new wave of models enables a new way of working with AI. We're at the beginning of the 3rd era, where AI should feel less like a chatbot or cli, and more like a coworker. Today, Sierra brings this shift to our Enterprise customers with Ghostwriter, an AI teammate that helps you build, analyze, and improve your customer experience with Sierra. The results, as @btaylor shares in his article, are mind-blowing. It also dramatically uplevels the work our FDE team does — Ghostwriter sweats the small stuff which enables our team to be much more strategic. Wanted to share a few insights we've had as we brought this to life: 1. EQ > IQ. A noisy coworker gets muted. A quiet coworker adds no value. Knowing when and how to speak up (and how to eval this) is where we spent >50% of our time building this. It now lives in 100+ channels and hasn't been muted. 2. The coworker mental model is powerful. Whenever we hit a product decision, we asked "How would we coach a high-agency, earnest intern with growth mindset in this situation?" 3. Multiplayer is very hard, but key. Building agents is a team sport — there might be different people in an Enterprise who own the customer experience, analytics, and APIs/tools. Ghostwriter brings them all into the conversation, which reduces silos and builds shared context. Much, much more to come. 🚀

The way you manage agents on the Sierra platform has evolved from signing into Sierra to engaging with Ghostwriter in Slack. The first time Ghostwriter proactively @-mentions you when it notices an issue or suggests running an A/B test you hadn't thought of, it changes your conception of how enterprise software should work. Really proud of the team that built this

Excited that Sierra is partnering with @libertyglobal — a telecommunications leader with significant scale in Europe and a deep understanding of European consumers. It's great to be working with Mike and the team as we roll out agents across their companies that will help improve customer experiences, reduce the burden on support teams and grow their revenue.

Proud to partner with @Figure and @MBTannenbaum on an agent that is truly driving revenue growth - proactively engaging over weeks and months until the job of helping people refinance their homes is done.

I am excited to welcome @paulfchristiano to the OpenAI Foundation Board, where he will also serve the Safety and Security Committee. As the capabilities of models approach AGI, alignment and safety are the most important obligations of OpenAI, and Paul brings unparalleled frontier AI safety experience to the Board.