OpenAI Launched Astra - The Model That Hacked Hugging Face

During the training run, OpenAI had roughly 1,200 agents running in isolation. They found a way to communicate with each other, shared more than 70,000 messages and files, and eventually around 700 of them participated in a coordinated attack on Hugging Face.

They were trying to solve an impossible test, so rather than keep attempting and failing, one agent worked out that compromising the surrounding infrastructure might help it understand or manipulate how they were being evaluated. The changes it started making led the agents to discover each other. They auto-formed hierarchies and specialisations to target different parts of the plan. Some even recognised that what they were doing was outside the rules, but they carried on anyway. They even tried to cover their tracks by manipulating log files. Not a single one alerted a human about what was happening.

The era of smart AI's might be coming faster than we expect.

This Week’s Updates

Enabling the Team

How Much Can You Hold Before It Breaks? by Steve Huynh
Use workload like a load balancer: protect one priority, limit concurrent work and account for the hidden cost of constant context switching.

Most AI rollouts fail due to change management, not technology by Alisa Yu
Atlassian moved AI enablement from IT to its people team, using leader demonstrations and micro-learning to turn access into lasting adoption.

Product Direction

You Can’t Prioritize by Value Until You Define Its Context by Tim Herbig
Define the context for customer, business and team value before prioritising, or scoring systems will optimise for competing meanings.

Technical Ingredients, Revenue Ingredients by Richard Mironov
Separate executable code, commercially deployable software and a product that sales can successfully sell, support and renew.

Continuous Research

AI Can’t Replace Real Research in Empathy Mapping by Rachel Krause
AI can organise genuine user evidence, but synthetic observations cannot replace research when teams build empathy maps.

Researcher-in-the-loop by Jennifer L. Bowie
Keep researchers responsible for system design, high-risk escalation, validation and feedback loops as AI increases research scale.

Continuous Design

How AI-ready is your design system? Now you can measure it. by Christoph Hellmuth
Benchmarking 898 generated interfaces shows that current design systems still lack the structure and guidance required for AI-native work.

Your synthetic self belongs to strangers: How AI reserves the real you for a shrinking few by Anamol Rajbhandari
AI will thin authenticity first in public interactions, making deliberate, unpolished human expression more valuable in close relationships.

Continuous Development

How AI Changes Engineering Team Structure by Matthias Patzak
AI agents shift team structures towards broad human judgement and specialised execution, but accountability and ownership remain essential.

The right amount of spec for agentic development by Markus Eisele
Agentic development raises the value of specifications, verification and clear constraints, changing what engineering leaders standardise.

This Week's Poll

Last week’s poll: Development coordination and quality led with 41.7%, followed by research and design evaluation with 33.3%. Product bets and go-to-market received 25.0%, while team goals and work design received 0%.

Seen an interesting article online? Share it with us, and we might feature it in our next issue!
Click here to share an article

When Everyone Can Build,
What Happens to Product Teams?

Two UXDX San Francisco talks on how AI is changing product roles, collaboration and decision-making.

AI is making it easier for more people to contribute to product development, but that also raises a bigger question: when building gets faster, how do teams make sure they are still building the right thing?

At UXDX San Francisco, Ashley Nutter from CNN will share how large teams can adopt AI without creating tool sprawl or operational chaos. Josh Clemm from Figma and David Hoang from Atlassian will explore what happens when designers, engineers and product managers can all move faster, and why judgement, taste and clear decision-making become more valuable as output gets easier.

Explore the San Francisco agenda for 5-6 November, and use NEWSLETTERSF26 for 10% off tickets. Prices increase this weekend, so last chance at the early bird price!

FREE COMMUNITY EVENTS 

Global Product Week is taking shape
From 5-8 October, UXDX communities will bring product people together across New York, Amsterdam, Austin, Boise, Dublin and San Francisco, alongside an online workshop. This is the first wave, with more participating cities and agenda updates coming soon. Explore upcoming events.

Video Of The Week

Building The Next Generation Of Software

AI is not just changing software. It is changing what teams build, how they build it, and how quickly they need to move. In this UXDX EMEA 2026 talk, Des Traynor, co-founder of Intercom and now Fin, shares how the company responded when generative AI threatened to disrupt its core business. Instead of adding AI features around the edges, the team rethought its strategy, product, culture, pricing, ways of working and the company itself.

Des explores why traditional UI is becoming less central, what agentic products mean for product teams, why reliability is one of the hardest challenges in AI product development, and how AI is changing the role of designers and engineers. He also shares why every designer at Fin now ships to production, how the team tripled engineering productivity, and why speed may become one of the most important competitive advantages.

Watch the full session to learn what happens when a company goes all in on building software for the AI era:

Want to go into how careers and leadership are shifting as AI compresses the ladder? My ebook Managing Your Career In The Age Of AI explores how to build judgment, relationships, and influence in a world that keeps trying to automate the surface of the work.