Have We Solved the AI Agent Oversight Issue?
Have We Solved the AI Agent Oversight Issue?
NVIDIA announced its Open Agent Safety Platform the morning we recorded. It's an open-source runtime (OpenShell) plus a hardware watchdog (Sentry) designed to limit what AI agents can do and quarantine them when they go beyond their boundaries. More than 100 organizations signed on, including Anthropic, SAP, Microsoft, IBM, and Salesforce.
So is agent oversight solved? The three of us don't agree.
🔹 Take 1: The conversation finally moved to the right place. A model alone can't act. It takes an agent, a program, or a person to carry out what it recommends. NVIDIA's announcement shifts the focus from the models to where the work actually happens. But enforcement still depends on NVIDIA hardware, and identity and access questions remain unresolved.
🔹 Take 2: Security isn't governance. An agent can stay inside its permission boundary and still cause harm. The platform doesn't answer who authorized the agent, for what purpose, under what constraints, or with what remedy when people are harmed. And a quiet warning for every business: check your insurance policies.
🔹 Take 3: The industry is regulating itself. No industry has successfully self-regulated. Not banking, not energy. Is this the wolf guarding the hen house?
We also cover NVIDIA's pending $12.9B Hugging Face acquisition, what the OpenAI–Hugging
Face breach showed us, research on agents tampering with their own monitoring traces, and why putting liability where AI is deployed may be the most effective control of all.
The bottom line: The question isn't whether an agent can escape the sandbox. It's whether we've put too much trust inside it in the first place.
đź“Ś Sources & further reading
🎙️ Three Takes on AI: three perspectives on the business of AI, every two weeks.Hosts: Brian Silverman, Campbell Robertson, Michael Muhlfelder
Comments