Five AI safety sessions every founder should have on their TechCrunch Disrupt 2026 agenda
Sep 22, 2026, 8:00 AM · TechCrunch
TechCrunch’s Disrupt agenda frames trust as the product gate — five safety sessions that treat agents, cloud, and robots as shipping risks, not keynotes.
Why it matters
AI founders aren’t debating hypotheticals anymore. Agents with system access, robots in unpredictable spaces, and autonomous vehicles all raise the same question TechCrunch puts up front: would you trust what you’re selling with real authority?
Disrupt 2026 (October 13–15, Moscone West) is packing five sessions across the AI Stage and Real World AI Stage around that question. The framing is commercial, not abstract — security, governance, and physical-world failure modes as blockers between a demo and a paying enterprise.
From the desk
We’re glad the conference is putting safety on the main stages instead of a side room. The lineup mixes the right kinds of pressure: Anthropic’s Cat de Jong on why some Claude deployments deliver and others stall in pilot for 18 months; Okta’s Ric Smith and NanoCo’s Gavriel Cohen on agent security when app-level permissions aren’t enough; AWS’s Rudy Mitra with Luta Security’s Katie Moussouris and Wendy Nather on cloud complexity as AI goes autonomous.
The Real World AI Stage sessions matter more to us. Shield AI’s Nathan Michael, GM’s Mikell Taylor, and Waabi’s Raquel Urtasun on systems where failure isn’t a bad autocomplete — it’s vehicles, aircraft, industrial gear. NVIDIA’s Les Karpas on the data gap that still keeps robots waiting for their ChatGPT moment. That’s the hard problem: language models scaled on internet text; physical AI still lacks comparable training pools.
Our read is blunt. Useful AI earns trust when founders treat security and safety culture as product requirements, not compliance theater. These sessions won’t invent that culture — but they name the buyers’ checklist. I’m watching whether the talk stays at “enterprise wants observability” or gets specific about permission models, kill switches, and what “safe enough” means when the system leaves the screen.
The conference expects 10,000+ attendees and 250+ speakers. Amid the pitch noise, these five rooms are where the industry admits the trust gap is the real go-to-market risk.
Context
Reporting from TechCrunch’s events desk, September 22, 2026 — a promotional agenda piece for Disrupt, not independent safety research. Speakers and session titles are as listed by TechCrunch.
Who feels it
- AI founders selling to enterprise
- Expect security, governance, and observability questions earlier in the sales cycle; pilots that never ship remain the default failure mode.
- Agent builders
- Application-level permissions are being called out as insufficient — infrastructure-level agent security is becoming a buyer requirement.
- Physical AI / robotics teams
- Data scarcity and validation standards — not just model cleverness — are the blockers to deployment trust.
- Enterprise buyers
- These sessions surface the checklist to demand: production evidence, permission architecture, and physical-world failure handling.
What to watch
- Whether Disrupt sessions produce concrete permission or kill-switch guidance, or stay high-level.
- Anthropic’s public takeaways from enterprise Claude deployments that actually leave pilot.
- Follow-on standards talk from Open Source Robotics Alliance / ROS community on GPU-aware data interfaces.
- Any regulatory or insurer language that echoes the “failure is not an option” framing for autonomy.