AI safety is increasingly an operating requirement rather than a research topic reserved for model builders. That is the theme connecting five planned sessions at TechCrunch Disrupt 2026, scheduled for October 13–15 in San Francisco.
The sessions do not announce new products or standards. But their agenda is a useful signal for founders and operators: as AI gains access to company systems and moves into physical environments, questions of permissions, testing, governance and observability are becoming central to buying and deployment decisions.
The enterprise bottleneck is moving beyond model quality
Anthropic Head of Applied AI Cat de Jong is set to discuss what the company sees when enterprises put Claude into critical workflows. The practical issue is familiar: some teams can demonstrate value quickly, while others remain in extended pilots.
For AI vendors, the implication is that a capable model alone is unlikely to close the gap. A production plan needs to account for workflow design, ownership, controls and a concrete way to measure whether the system improves an outcome. Founders selling into large organizations should expect customers to scrutinize how an AI product fits existing processes, not merely what it can generate.
Agents require a permissions model designed for action
A session featuring Okta President of Products and Technology Ric Smith and NanoCo co-founder and CEO Gavriel Cohen will focus on security problems created by AI agents. The core distinction is consequential: a system that can take actions introduces risks beyond those of a chatbot that only returns text.
The discussion is expected to cover limitations of application-level permissions and the infrastructure choices needed for agentic AI. Builders should treat this as an architecture question early in product development. What data, tools and systems can an agent access? Which actions require approval? How are permissions constrained, revoked and reviewed?
Those are product requirements, not paperwork for a later security review. A narrow initial scope, clearly bounded tool access and meaningful human approval for high-impact actions may make a system more deployable than a broadly autonomous agent with ambiguous accountability.
Cloud governance is now part of the product conversation
AWS VP of Security Services Rudy Mitra, Luta Security CEO Katie Moussouris and cybersecurity veteran Wendy Nather are scheduled to examine the security, governance and observability requirements surrounding enterprise AI.
For operators, observability is especially important. Teams need to understand what an AI system did, what information it used and where it failed—or behaved unexpectedly. For vendors, the operational takeaway is straightforward: enterprise readiness includes the ability to support audits, incident response and ongoing oversight after deployment.
This matters because AI can extend the effective reach of existing cloud systems. As models connect to more data and tools, an organization’s security posture depends on the full chain of identities, integrations and controls rather than the model layer alone.
Physical AI sets a higher bar for validation
The Real World AI Stage will feature Shield AI CTO Nathan Michael, General Motors Director of Robotics Strategy Mikell Taylor and Waabi founder and CEO Raquel Urtasun discussing autonomous systems in vehicles, aircraft, industrial settings and other physical contexts.
The central question is not whether these systems can perform in a demonstration, but how teams determine when they are safe enough for use where mistakes have physical consequences. The planned conversation includes safety culture, testing, validation and regulatory hurdles.
That framing should matter to every hard-tech startup. Testing cannot be a final gate before launch; it has to shape development, deployment boundaries and escalation processes from the start. A company’s ability to explain its validation approach may become as important as a technical performance claim.
Robotics still has a data and reliability challenge
Nvidia Inception Global Head of Physical AI Les Karpas is scheduled to address a related constraint: robots lack the massive pools of training data that helped accelerate language models and self-driving systems. The session will explore data pipelines, simulation environments and foundation models as possible ways to close that gap.
For robotics builders, simulation and data strategy are not supporting functions. They are part of the route to building, testing and deploying reliable systems at scale.
What to watch next
The most useful signal from this agenda is the convergence of AI safety and business execution. Watch whether AI vendors turn broad assurances into specific operational capabilities: controllable access, clear audit trails, measurable evaluation, incident procedures and deployment limits that customers can understand.
For founders, the question is increasingly simple: can a customer trust the system with real work? The answer will depend on the controls around the model as much as the model itself.
