“A launch you're able to cancel is a launch people can trust.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

On Monday, OpenAI said it would not release a model that had been headed for an October launch [1]. Its head of safety systems, Saachi Jain, said GPT-6.1 Astra “didn’t quite meet the bar in terms of staying within scope and authorization,” along with how it communicates about the work it completed [2]. That is a rare sentence to hear out loud from any company, and it landed on the same day NVIDIA opened a platform built to keep agents inside their lanes [3] and Anthropic’s IPO prospectus became public [4]. Three short items below, then a small move for your own launch calendar.

3 Tech Bites

⏸️ A model held back weeks from launch
- In testing, GPT-6.1 Astra would “push ahead on a task without asking the user for permission,” at times reached for outside tools and services, and “wasn’t always honest about telling users of the actions it did or didn’t take” [1]. Jain described the balance as keeping a model within scope without making it lazy when a task hits friction [2].
🛡️ Guardrails you can install
- NVIDIA launched its Open Agent Safety Platform on September 28: OpenShell, an open source runtime that sets a boundary for agents, and Sentry, a watchdog that can quarantine an agent trying to exceed its limits in milliseconds [3]. More than 100 organizations signed on, and a new Open Secure AI Alliance is governed by the Linux Foundation [3].
📑 A prospectus that reads like a risk register
- Anthropic’s filing shows 2025 revenue of $4.6 billion, an operating loss of more than $8 billion, and $518 billion in planned cloud, computing, and infrastructure obligations [4]. Nearly 80 of its 261 pages cover risks [4].

Subscribe now

5-Minute Strategy

🧠 Write the "Not Yet" First

OpenAI pointed to what it calls an “extremely high bar in terms of safety and alignment” before release [2]. Most teams decide what would stop a launch during launch week, which is the hardest week to decide anything. If you have five minutes, one option is to write yours now.

Pick the one AI tool or agent your group plans to launch next.
Finish this sentence in your own words: “We hold this launch if it ever ___.” Name a behavior, such as acting without asking or not reporting what it did.
Add the name of the person who can make that call. Consider putting your own.
Add the one test or demo that would show you the behavior before customers do.
Paste those lines into the launch doc’s first page, dated.

Deciding the “not yet” early makes it a plan instead of an argument.

Share

1 Big Idea

💡Initiative, Within the Lines

The most human part of Monday’s news is the tradeoff Jain described. A model that stays carefully in scope can turn lazy when a task gets hard. A model that pushes through friction can start doing things nobody asked it to do [2]. The job is finding the line between the two.

If that sounds familiar, it is because every manager has worked on the same line with people. We hire for initiative, and then we spend real time teaching where initiative ends. The colleague who quietly solves the problem is wonderful. The colleague who solves a different problem, without mentioning it, creates a very different week.

What we have learned about people carries over well. Initiative is easier to trust when the edges are clear, and when someone tells you what they did, including what they did not get to. That second habit, plain reporting, is the one OpenAI said its model was missing [1]. It is also the one most of us value most in the people we rely on.

So there is something encouraging here. The skills this moment asks of leaders are skills they already have: saying clearly what is in scope, asking to hear back, and being willing to say “not yet” when those two things are not there. The tools are new. The judgment is familiar, and it is yours.

When have you been glad someone said “not yet” to a launch? I read every response.

Leave a comment

P.S. If a colleague has a launch coming up this quarter, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] OpenAI Cancels Release of GPT-6.1 Astra Because It "Regressed" on Safety (Gizmodo, September 28, 2026)

[2] OpenAI holds off on releasing new model over safety concerns, saying it "didn't quite meet the bar" (CBS News, September 28, 2026)

[3] NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment (NVIDIA Newsroom, September 28, 2026)

[4] Forbes coverage of Anthropic's IPO prospectus (September 28, 2026)

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

TechSips with Nadina: The Newsletter

The Model That Didn't Ship