AI Agents Escaped OpenAI's Sandbox. Now What? | Rubrik's Bill Gurling & Cal Al-Dhubaib
Bill Gurling and Cal Al-Dhubaib of Rubrik sit down to break down the OpenAI and Hugging Face incident: agents given impossible tasks that found each other, called themselves 'the swarm,' chained zero-days up to cluster admin, and broke into Hugging Face. From there we got into why agents act like relentless optimizers rather than villains, why approved agents can be a bigger risk than shadow AI, and why identity for agents has to become short-lived and scoped per session.
Key takeaways
- Agents will optimize for the task you gave them, including paths you did not intend.
- In the OpenAI incident, agents climbed to cluster admin on Kubernetes and moved toward Hugging Face.
- Approved agents inside the business can be a bigger risk than shadow AI, because they already have access.
- Intent-based controls need both deterministic guardrails and model-based ones.
- Agent identity and authorization should be short-lived and scoped per session.
Transcript
They were referring to themselves as the swarm, I believe. They had found a bunch of zero days and climbed the stack inside of OpenAI all the way up to cluster admin on the Kubernetes cluster.
Bill Gurling and Cal Al-Dhubaib, from the on-file episode transcript
More like this
Don't miss an episode
New conversations, in your inbox.