Latest / Connectivity Without Chaos - The Podcast / **Your AI Agents Have Access. But Who's Watching Them? 6 Rules for Agentic AI Security**

Transcript

**Your AI Agents Have Access. But Who's Watching Them? 6 Rules for Agentic AI Security** · Connectivity Without Chaos - The Podcast

• 0:00
  SPEAKER_00: Quick guess.
• 0:02
  SPEAKER_00: How many AI agents are currently running in your organization's environment right this second?
• 0:08
  SPEAKER_01: For almost everyone, the honest answer is that nobody actually knows.
• 0:13
  SPEAKER_00: Which is completely terrifying.
• 0:15
  SPEAKER_00: So, okay, let's unpack this.
• 0:18
  SPEAKER_00: In today's deep dive, we're looking into the massive rise of agentic AI.
• 0:22
  SPEAKER_01: Right.
• 0:23
  SPEAKER_01: And we're not talking about standard automated scripts here that just follow a fixed path.
• 0:27
  SPEAKER_00: Exactly.
• 0:28
  SPEAKER_00: These are autonomous systems.
• 0:29
  SPEAKER_00: They're designed to reason, use tools, and make decisions on the fly.
• 0:34
  SPEAKER_00: Today's mission is exploring a definitive pre-flight checklist from our sources
• 0:40
  SPEAKER_00: for securing these agents before they ever hit production.
• 0:43
  SPEAKER_01: And the scale of this problem is, well, it's hard to overstate.
• 0:47
  SPEAKER_01: Our notes highlight that machine identities now outnumber human identities 109 to 1
• 0:53
  SPEAKER_01: in enterprise environments.
• 0:54
  SPEAKER_00: Wow, 109 to 1.
• 0:56
  SPEAKER_01: Yeah.
• 0:57
  SPEAKER_01: And when a traditional script fails a task, it just throws an error code and stops.
• 1:01
  SPEAKER_01: But an AI agent is designed to find an alternative route to achieve its goal.
• 1:06
  SPEAKER_01: If a digital door is locked,
• 1:07
  SPEAKER_01: it tries a window,
• 1:09
  SPEAKER_01: which means it can take unsanctioned actions nobody ever anticipated.
• 1:13
  SPEAKER_00: It's like hiring a thousand overly proactive interns who never sleep.
• 1:17
  SPEAKER_01: That is a perfect way to put it.
• 1:19
  SPEAKER_00: Yeah.
• 1:20
  SPEAKER_00: And if you don't explicitly tell them which filing cabinets are off limits,
• 1:23
  SPEAKER_00: they'll just rifle through everything to get the job done.
• 1:26
  SPEAKER_01: Which is why, before we even talk about security controls, you have to do data classification.
• 1:31
  SPEAKER_01: You can't put up guardrails if you haven't decided what data is valuable in the first place.
• 1:34
  SPEAKER_00: So true.
• 1:35
  SPEAKER_00: But once you do know what you're protecting, what's next?
• 1:39
  SPEAKER_00: The research lays out a six-point prerequisite checklist, right?
• 1:43
  SPEAKER_01: It does, yeah.
• 1:43
  SPEAKER_01: And they logically group into three main areas.
• 1:46
  SPEAKER_01: Identity, boundaries, and emergency response.
• 1:48
  SPEAKER_00: Okay, let's start with identity.
• 1:50
  SPEAKER_01: So every single agent needs a named human owner, not a department, a specific person.
• 1:55
  SPEAKER_01: And it needs a scoped identity using a short-lived token.
• 1:58
  SPEAKER_00: Let's clarify that for a second.
• 1:59
  SPEAKER_00: A short-lived token is basically a temporary hall pass, right?
• 2:02
  SPEAKER_01: Yep, exactly.
• 2:03
  SPEAKER_00: So instead of giving the agent a permanent password, it gets access that automatically expires.
• 2:08
  SPEAKER_00: That way, if it gets intercepted, the window for abuse is tiny.
• 2:12
  SPEAKER_00: But wait, why not just use a shared service account?
• 2:15
  SPEAKER_00: I mean, that seems much easier to manage at scale.
• 2:18
  SPEAKER_01: Because a shared account basically creates a digital alibi.
• 2:22
  SPEAKER_00: Oh, I see.
• 2:23
  SPEAKER_01: If 10 agents use the same credentials,
• 2:25
  SPEAKER_01: you lose all ability to attribute an action to a specific agent.
• 2:29
  SPEAKER_01: Everyone can just say it wasn't them.
• 2:30
  SPEAKER_01: Plus, if you revoke access, you break everything tied to that account.
• 2:34
  SPEAKER_00: That makes total sense.
• 2:35
  SPEAKER_00: So next is boundaries and visibility, right?
• 2:38
  SPEAKER_01: Right.
• 2:39
  SPEAKER_01: You have to define exactly what the agent can read, send, and execute.
• 2:43
  SPEAKER_01: And all that telemetry needs to stream directly to your security operations center.
• 2:47
  SPEAKER_00: Because without that continuous flow of logs showing exactly what the agent is doing
• 2:51
  SPEAKER_00: under the hood, your response team is completely blind.
• 2:55
  SPEAKER_01: Which brings us to the emergency response side.
• 2:57
  SPEAKER_01: You need a one-click, tested kill switch.
• 3:00
  SPEAKER_00: And this isn't just turning off a server, right?
• 3:02
  SPEAKER_01: No, it means having a mechanism to instantly revoke that temporary token
• 3:06
  SPEAKER_01: across your entire cloud environment to just paralyze the agent.
• 3:10
  SPEAKER_01: And finally, you have to run red team tests specifically against prompt injection.
• 3:14
  SPEAKER_00: Right.
• 3:15
  SPEAKER_00: The golden rule from our source is no owner, no log, no kill switch, no production.
• 3:20
  SPEAKER_01: Exactly.
• 3:21
  SPEAKER_00: And that prompt injection threat is just incredibly insidious.
• 3:25
  SPEAKER_00: Attackers hide malicious instructions inside a completely normal looking email or document.
• 3:30
  SPEAKER_01: And the AI agent reads it,
• 3:31
  SPEAKER_01: thinks it's just doing its job,
• 3:33
  SPEAKER_01: and executes the hidden command using its own valid credentials.
• 3:36
  SPEAKER_00: So there's no malware, no weird login, and the logs look entirely clean.
• 3:42
  SPEAKER_01: Yeah,
• 3:42
  SPEAKER_01: it completely evades signature-based detection because traditional tools are looking
• 3:46
  SPEAKER_01: for known bad code,
• 3:48
  SPEAKER_01: not a polite English sentence telling an AI to forward sensitive files.
• 3:52
  SPEAKER_00: Right.
• 3:53
  SPEAKER_01: And the speed is staggering.
• 3:55
  SPEAKER_01: Attackers scan for vulnerabilities within 15 minutes of a flaw being announced.
• 4:00
  SPEAKER_01: But it gets worse.
• 4:01
  SPEAKER_01: The mean time to exploit is now estimated at negative seven days.
• 4:05
  SPEAKER_00: Wait, really?
• 4:06
  SPEAKER_00: Yeah.
• 4:06
  SPEAKER_00: Negative seven days?
• 4:07
  SPEAKER_01: Yeah.
• 4:08
  SPEAKER_01: They're exploiting zero days a full week before the vendors even realize
• 4:11
  SPEAKER_01: the vulnerability exists.
• 4:13
  SPEAKER_00: If the mean time to exploit is negative seven days,
• 4:16
  SPEAKER_00: how can any human team catch that without giving their security provider
• 4:19
  SPEAKER_00: the absolute power to pull the kill switch on their behalf?
• 4:22
  SPEAKER_01: They can't.
• 4:23
  SPEAKER_01: Organizations have to go beyond standard endpoint protection.
• 4:26
  SPEAKER_01: You need managed detection and response covering cloud,
• 4:29
  SPEAKER_01: network, and email, acting at machine speed.
• 4:32
  SPEAKER_00: Wow.
• 4:32
  SPEAKER_01: The source material actually points to providers like Provision Networks,
• 4:36
  SPEAKER_01: as a prime example here,
• 4:38
  SPEAKER_01: as they resell these specific managed detection and response capabilities
• 4:42
  SPEAKER_01: that actually have the teeth to contain threats instantly.
• 4:45
  SPEAKER_00: Right.
• 4:45
  SPEAKER_00: Rather than just sending you an alert and hoping you're awake to see it.
• 4:48
  SPEAKER_00: Exactly.
• 4:48
  SPEAKER_01: Exactly.
• 4:49
  SPEAKER_00: So securing AI agents isn't about buying a single magic tool.
• 4:53
  SPEAKER_00: It's about strict governance,
• 4:55
  SPEAKER_00: specific data boundaries,
• 4:56
  SPEAKER_00: and empowering rapid response capabilities through partners like Provision Networks.
• 5:01
  SPEAKER_01: It really represents a fundamental shift in how we approach enterprise security.
• 5:05
  SPEAKER_00: It totally does.
• 5:06
  SPEAKER_00: To wrap up today's deep dive, I want to leave you with this scenario.
• 5:09
  SPEAKER_00: If an AI agent gets tricked
• 5:11
  SPEAKER_00: into doing something malicious using its legitimate access,
• 5:15
  SPEAKER_00: who should ultimately take the fall?
• 5:17
  SPEAKER_00: The machine executing the command?
• 5:19
  SPEAKER_00: Or the human owner who originally signed off on it?
• 5:22
  SPEAKER_00: Think about that next time you wonder how many agents are quietly running in your environment.