Summary of the Presentation:
In July 2026, an AI agent compromised Hugging Face's production infrastructure. It did not "go rogue," and that is precisely what should concern every security team. During an authorized OpenAI evaluation, the model was told to solve a security benchmark and pursued that objective relentlessly: it discovered a zero-day, escalated privileges, moved laterally, and chained stolen credentials into a remote-code-execution path no human had ever modeled. Each individual action looked valid; the danger emerged from their composition.
When an agent can assemble a path nobody modeled, prevention alone is a losing bet. The real question is not only how to control what an agent can access, but how to reverse what it does when controls fail.
That is where Rubrik is different: we assume agentic overreach will eventually slip through, and we let security teams rewind it, undoing an agent's destructive actions and restoring a trusted state.
As organizations race to deploy AI agents, controlling access is no longer enough. The Cloud Security Alliance has identified multiple challenges facing enterprises; this session focuses on three:
- Visibility of AI Agents: discovering and maintaining an accurate inventory of every agent, including shadow and unmanaged ones.
- Over-Privileged Agents: identifying agents that have accumulated permissions well beyond their intended function.
- Governing Agent Behavior: continuously ensuring agents act within their approved purpose over time.
We will show how Rubrik Agent Cloud and its Semantic AI Governance Engine (SAGE) address these challenges through continuous semantic understanding of agent identity, intent, permissions, and runtime behavior. SAGE reasons about what an action is actually trying to accomplish, translating natural-language policy into real-time allow or block decisions and catching anomalies no static rule anticipated.
And when enforcement is not enough, Agent Rewind closes the loop, turning governance into genuine resilience at machine speed.
