“The two key fundamentals that I've identified is the persistence or ephemeral nature of the identity and the deterministic or non-deterministic nature of the identity. … Humans, we are non-deterministic, but we're also persistent. … Services, machine identities, they are ephemeral. They pop in and out of existence, but they're deterministic. … Agentic identities are the worst of both. They're both ephemeral and non-deterministic.”
sputnikintelligence.com
The host asked what changes about authorization when agents, not people, do the work. Weis argues agents shouldn't be treated as just another employee or another service account, and builds the case from two properties of an identity.
…an intuition that requires adjusting. And the more I'm catching myself doing things that I shouldn't and actually delegating to AI, I find that it completely changes my experience. But overall, it enables me to kind of go with the flow, go with the times, and don't feel so overwhelmed. I think.
…from yourself, a software engineer, your persona, right, your archetype, into agents. So, so, particularly about talking about authorization, what has changed, or what is changing? Because things have not settled yet with regards to authorization with agentic driven with agents running the show.
…break here. Thinking about identities in general, that has changed a matter. So, we have human identities, those we've had for a while, and we recently added machine identities like services. And now we're adding AI Agents. And one of the key questions there is. Are AI Agents the same as humans or as services? And the way I like to break this down is try to boil it down to the fundamentals. So the two key fundamentals that I've identified is the persistence or ephemeral nature of the identity and the deterministic or non-deterministic nature of the identity. So if you put that down, if we look at humans, we are non-deterministic. but we're also persistent. We're not ephemeral. We don't pop in and out of existence. We don't create multiple copies of ourselves, etc. Services, machine identities, they are ephemeral. They pop in and out of existence, but they're deterministic. They'll continue to behave the same way, basically at an infinitum in a very…
I've heard of all these problems. I've already heard about making the humans developing it, not only the software engineers, let's say, anthropic, as an example, any AI provider, but the business, the management responsible for the consequences of them running wild or not being legally liable, the…
Yeah, the confused deputy problem is actually a classic security problem. The idea there is that you have a deputy that you deputize to do things and it's well-intentioned. It wants to do the right thing, but it can get confused. It's trying to do what you wanted, but it does something slightly off…
Yeah, I see. Thanks for that explanation. So, you've touched upon before standards. I might be biased because I'm a contributor and a part of the review committee of the model context protocol. But I do think genuinely that it has become the de facto standard for tool calling. Right? I do think so.…