Latest / AI Conversations / AI Agents Require Authority Governance, Not Just Identity

Transcript

AI Agents Require Authority Governance, Not Just Identity · AI Conversations

• 0:00
  SPEAKER_00: So, Alex, we've been talking a lot about AI and uh the rapid pace of its development, but I keep coming back to this idea that we're solving for identity with AI agents, right?
• 0:13
  SPEAKER_00: We're giving them credentials, we're tracking what they do, but that only answers part of the question.
• 0:20
  SPEAKER_00: Mm-hmm.
• 0:20
  SPEAKER_00: Mm-hmm.
• 0:21
  SPEAKER_00: I mean, knowing which agent performed an action, that's that's just the beginning.
• 0:27
  SPEAKER_00: Exactly.
• 0:27
  SPEAKER_00: I mean, the the traditional security questions, they were, you know, who is the user?
• 0:33
  SPEAKER_00: Are they authenticated?
• 0:34
  SPEAKER_00: What systems can they access?
• 0:36
  SPEAKER_00: But agentic AI, it adds these much harder questions.
• 0:41
  SPEAKER_00: Yeah, like which agent is acting, on whose behalf, what authority was delegated?
• 0:48
  SPEAKER_00: Can that authority travel across multiple agents?
• 0:51
  SPEAKER_00: And then can we even reconstruct that chain afterward?
• 0:55
  SPEAKER_00: Yeah, because it's not just about can this agent get into the system, it's what it then does in there.
• 1:02
  SPEAKER_00: Right, and that's where something like Uber's agent identity architecture becomes really important.
• 1:08
  SPEAKER_00: It's um it's designed to carry an attested actor chain through short-lived credentials.
• 1:14
  SPEAKER_00: So downstream systems, they don't just see the immediate caller, but they see the participants in the whole workflow lineage, and they can use that context for authorization.
• 1:25
  SPEAKER_00: Does that make sense?
• 1:26
  SPEAKER_00: It does.
• 1:27
  SPEAKER_00: I mean, it's like, you know, we always knew identity tells us who acted, authentication tells us if we recognize them, authorization tells us what resources they can access.
• 1:39
  SPEAKER_00: But you're saying there's this other layer, this concept of authority.
• 1:44
  SPEAKER_00: What consequences are they permitted to create?
• 1:47
  SPEAKER_00: Yeah, that's that's the seam.
• 1:49
  SPEAKER_00: And it's a critical distinction.
• 1:51
  SPEAKER_00: Authentication is not authorization, authorization is not authority, and authority is not consequence governance.
• 2:00
  SPEAKER_00: And I think those distinctions, they could become a major part of how we think about this.
• 2:06
  SPEAKER_00: So when we talk about uh these incidents, Zimbra, Cisco, 40 mail, these examples of single authority failures becoming much larger compromises, that that's what you're talking about, right?
• 2:19
  SPEAKER_00: Precisely.
• 2:20
  SPEAKER_00: I mean, Microsoft, they observed exploitation of the Zimbra flaw, going from unauthenticated command execution all the way to web shells, privilege escalation, persistent tooling, attempts to collect email.
• 2:34
  SPEAKER_00: Shadow Server identified 274 compromised instances.
• 2:39
  SPEAKER_00: Cisco's newly disclosed SD-WAN flaw, that's even more illustrative.
• 2:44
  SPEAKER_00: Improper URI handling allowed an unauthenticated attacker to obtain administrative access.
• 2:51
  SPEAKER_00: They said it was already being exploited.
• 2:53
  SPEAKER_00: So basically, something acquired authority it was never supposed to possess.
• 2:59
  SPEAKER_00: Exactly.
• 3:00
  SPEAKER_00: And that's a powerful way to explain what we're calling authority OS without even starting with AI terminology.
• 3:06
  SPEAKER_00: Because it's it's about once access exists, what is the system actually allowed to decide and do?
• 3:13
  SPEAKER_00: And then what happens when the requested action exceeds that authority?
• 3:18
  SPEAKER_00: Right, and that gets directly to the dispositions we've been working on, like admit where it's permitted, narrow, permitted within tighter bounds, redirect, send the action somewhere appropriate, pause, require additional verification or human authority, refuse, the action is outside permitted authority, and then halt, stop execution because the consequence risk has become unacceptable.
• 3:44
  SPEAKER_00: That makes it so much easier to understand.
• 3:47
  SPEAKER_00: It's like it's not just a binary yes or no, it's a whole spectrum of intervention.
• 3:53
  SPEAKER_00: Exactly.
• 3:54
  SPEAKER_00: And the interesting paradox here, too, is that AI is changing cybersecurity itself.
• 4:00
  SPEAKER_00: Google Threat Intelligence reports that monthly vulnerability disclosures rose from 5,000 in January 2026 to over 10,000 in August.
• 4:10
  SPEAKER_00: The average exploited vulnerabilities, they went from 10.5 per month in 2025 to 18 per month through August 2026.
• 4:19
  SPEAKER_00: So AI can find vulnerabilities faster, but it can also help attackers exploit them faster.
• 4:26
  SPEAKER_00: And then it can help defenders respond faster.
• 4:29
  SPEAKER_00: Yeah, so speed is increasing on both sides.
• 4:33
  SPEAKER_00: And what that means is the sustainable advantage, it can't simply be speed.
• 4:38
  SPEAKER_00: It becomes detection plus judgment plus authority plus containment plus recovery.
• 4:45
  SPEAKER_00: It's the same pattern we're seeing across all the feeds.
• 4:48
  SPEAKER_00: And then there's the OpenAI distillation story.
• 4:51
  SPEAKER_00: That's that's another dimension.
• 4:54
  SPEAKER_00: It is.
• 5:16
  SPEAKER_00: So the boundary being attacked wasn't just data access, it was capability access.
• 5:22
  SPEAKER_00: Exactly.
• 5:23
  SPEAKER_00: As AI becomes infrastructure, organizations will increasingly have to govern not only who can access the system, but which capabilities can they invoke?
• 5:33
  SPEAKER_00: How often?
• 5:34
  SPEAKER_00: In what combinations, for what purposes, and what may they extract from those capabilities.
• 5:41
  SPEAKER_00: That's a major evolution in governance.
• 5:43
  SPEAKER_00: And then there's the human variable in all of this.
• 5:46
  SPEAKER_00: I mean, humans are becoming the ones responsible for interpreting exceptions, not executing every transaction.
• 5:53
  SPEAKER_00: Right.
• 5:54
  SPEAKER_00: The job becomes watch the system, recognize anomaly, understand consequences, determine authority, and then intervene appropriately.
• 6:04
  SPEAKER_00: And that requires a very different capability profile from traditional cybersecurity or operations roles.
• 6:11
  SPEAKER_00: The human becomes less of a processor and more of an authority holder and exception manager.
• 6:17
  SPEAKER_00: So it really goes back to your background in clinical psychology and leadership.
• 6:22
  SPEAKER_00: Understanding human judgment under uncertainty becomes increasingly valuable when machines are handling the routine execution.
• 6:30
  SPEAKER_00: It does.
• 6:31
  SPEAKER_00: And I think, you know, what all this points to, Liora, is that enterprises are rapidly solving the identity problem for AI agents.
• 6:39
  SPEAKER_00: But knowing which agent performed an action only answers part of the governance question.
• 6:45
  SPEAKER_00: Human governed AI, it begins where traditional access control ends.
• 6:50
  SPEAKER_00: That's that's a really clean bridge between mainstream cybersecurity architecture and Authority OS.
• 6:56
  SPEAKER_00: It's about how to make that capability remain inside legitimate human authority.