Agent Breaches Are an Identity Problem — Inside Entro Security's Playbook After SailPoint
Key points
Key takeaways from a theCUBE interview with Entro Security co-founder and CEO Itzik Alvas, recorded live at SailPoint Navigate 2026 (October 2026):
Founded out of lived breach experience. Before launching Entro in 2023, Alvas led internal security at Microsoft and dealt with breaches rooted in non-human identities, while co-founder Adam Cheriki's employer suffered similar incidents; the pair started Entro to stop future organisations being breached through machine identities, credentials and permissions.
Why SailPoint. Entro is now part of SailPoint after the acquisition closed in June 2026, following a partnership that grew out of customer demand; Alvas argues a standalone non-human-identity solution forces customers to run between two platforms, and that identity governance for agent ecosystems is already past the tipping point where traditional tools can cope alone.
Permissions follow the person, not just the agent. Because an AI agent acts on behalf of a human, Entro correlates human and non-human identities so that one policy governs access whether it is used through a browser or an agent — a finance colleague querying OneDrive for a CEO's salary is legitimate, while the same prompt from outside finance is denied.
Three enforcement points for agent actions. Risk sits at the prompt, in the agent's plan and at execution: Alvas described a customer whose policy allowed read/write on OneDrive, where the agent planned to delete existing files and recreate them — an action blocked because deletion fell outside the permitted scope.
Breaches are an access problem, not a model problem. Most headline agent incidents trace back to an exposed credential an agent found and used; if an identity lacks permission, the attempt simply fails, so the priority is guardrails rather than judging whether behaviour was malicious or a mistake.
Permission ceilings stop privilege escalation. In agent swarms, a sub-agent should never hold more permission than the agent that spawned it, preserving the chain of delegation back to the original human instruction.
A fast path to NHI visibility. As a former CISO who was himself breached, Alvas says onboarding is now a matter of a few clicks: an inventory of machine accounts and agents, their permissions, owners, risks and gaps, plus real-time monitoring that can stop suspicious invocations in their tracks.
Agent breaches as identity news. Citing disclosures including Google's Gemini autonomously breaching companies, and similar announcements from Anthropic, OpenAI and Meta, Alvas frames them as identity failures — a username, password or machine account left findable in the wild — rather than model-alignment problems.
Read more: theCUBE on YouTube
