“Today the average user, the average app, has 95% of the permissions granted go unused. And we've gotten away with that because humans are ethical and wise and kind of slow, and applications are wired to do exactly one thing exactly one way. But an agent is none of these things. It's not wise, ethical or slow, and it is not a single path. So now all of these excess permissions come back to bite us.”
sputnikintelligence.com
The host asked how organizations should think about the risk from agentic deployments. The guest argues that over-permissioned environments were tolerable when only people and fixed-code apps used them, and recommends scoping agents to exactly what they need. The 95% figure is his own claim.
…And so then we see things like Hugging Face or whatever, where like this, in the goal of trying to reach the objective, things, surprising things happen because we can discover resources, we can discover the secrets to get to more resources, and we're not inhibited by time or ethics or wisdom.
So given that reality, how should organizations be considering their risk?
I would not be the first person to say this. I think the risks are substantial, right? There is really substantial risk to doing agentic deployments. And I think there's reasonable mitigations for those risks, but the risks of an agent doing something that is a surprise to you using resources you weren't aware that it had access to is, and that's even in the benign case. I'm not talking about malicious usage, right? Is very real. The thing that is mundane about all of this is that if we think, okay, an agentic application is just an application that uses LLM to make decisions, right? Uses an LLM to make decisions, then in that world, I think you it helps you to realize like, okay, well, it's not that different from another application. Again, it needs access to secrets, it needs permissions, it needs, you know, it needs to be invoked, right? So controlling those aspects, and I think most importantly, understanding what can an agent get to. The analogy I like to use is like, think about, you know, sort of too much cotton candy. I have a nephew who's 11 years old, right? I can imagine taking him out to the state fair and like giving him way too much caffeine and cotton candy and…
What sort of things are you seeing attackers doing in terms of using agentic technologies against identity systems?
We have, I say there's two kind of classes. One is the use of Agentic technology to inspect systems for vulnerabilities that can be exploited through conventional means, right? So I don't necessarily use the agent to do the Attack, although that might chain, but just the ability to inspect an…
That's Alex Weinert, Chief Product Officer at Semprus. The cybersecurity challenges shaping our future demand more than conversation. They demand action. On November 18th, the Aspen Cyber Summit brings together leading voices from government, industry, and civil society in Washington, D.C. to…