AI Incident Responses: Okta CSO On What Midmarket IT Leaders Should Do In The First Hour
Shutting down a compromised AI agent can destroy evidence. Okta’s Matt Immler explains what midmarket IT teams should do first in an AI incident.
There is often a panic-driven itch to switch off an AI agent the moment an incident happens, an immediate response to a problem that can spread across company systems within minutes.
As organizations put more AI agents into production, Matt Immler, regional CSO for the Eastern Americas at Okta, argues that pulling the plug may not be the ideal first move.
In an interview with MES Computing, Immler explains how IT teams, including those that rely on managed service providers, can cut off a compromised agent’s access while preserving the evidence they need, and why tracing everything the agent touched matters before an incident is considered contained.
Revoke The Agent’s Access Before Shutting It Down
According to Immler, IT teams still need to detect, contain and remediate a threat during an AI incident, as they would in any other type of cyber incident, but the pace at which an agent operates changes how those steps play out.
"You’re not dealing with the traditional actor who’s probably a person or a legacy tool. You’re dealing with something that can act and move a lot faster, so the methodology remains relevant, but the way in which you achieve it changes," he said.
Under that kind of pressure, many teams are tempted to shut a compromised agent down at once, a reflex Immler warned could be counterproductive.
“If you just kill it off or immediately shut it down, you risk losing important information about what was going on that’s probably going to help with the investigation later,” he said.
He favors cutting off the agent’s ability to act, which is what the industry’s talk of AI kill switches comes down to — giving administrators a way of “preventing it from creating new token requests, preventing it from any future authorizations.” Containing the agent this way preserves the evidence, because “you never know what’s sitting there in memory.”
Organizations without a commercial kill switch can also achieve the same result manually, as long as they have what Immler refers to as “a good accounting of what it is your agent has access to and where it’s deriving its permissions from."
“That way, just like in any directory or any rogue user, you can go in and turn off the access in its account.”
A team that knows where an agent is running may also be able to separate it at the network layer, “orphaning it in its own sandbox," he added.
For midmarket firms that depend on managed service providers, Immler believes the question is whether incident monitoring stops at an alert.
“Is it just looking to alert so that it can say that something’s going on and then waiting for a team to actually do something manually? In which case you’re losing valuable time in responding to an incident,” he said. Programming the platform to trigger a universal logout can help close that gap, with the MSP signing off on the action while root access stays with the organization.
Put Every Agent In The Directory
Knowing what each agent can reach depends on the organization knowing the agent exists in the first place. Immler cautioned against giving in to the pressure to roll out AI at speed, which he believes leaves organizations with a pile of technical and security debt to clear later, often at the point when an incident has already exposed the gaps.
Agents are non-deterministic and, in his words, “very human-like in the way they act,” which sets them apart from a typical service account that behaves the same way every time. For that reason, he argues they should be treated almost as if they were human users and held to the same governance as the people working alongside them. That means each agent “should have its own profile in the directory, managers, governance requirements, reviews of access, and it should be accounted for in the actual environment,” Immler said.
The goal is to bring each agent into the directory from the moment it is deployed, rather than leaving IT teams to search for agents already running in their environment after the fact.
Organizations that already have agents deployed without a complete inventory still have options, as observability tools can now help IT teams “identify AI agents within their infrastructure” and bring them under the right governance processes," he said.
Inside The First Hour Of An AI Incident
Once the agent has been isolated and its access revoked, the rest of the first hour should focus on determining whether customer data or critical systems were exposed and what may have been affected.
From there, Immler said investigators should determine:
- Who initiated the request?
- Which agent acted?
- What permissions did the agent have at the time?
- What other tools or agents did it call?
- Did the request originate internally?
- Did it come from an external prompt?
- Did an automated task cause the agent to behave in a way that was not intended?
Another important task is following what the agent set in motion, since it may have called other agents or been given additional permissions, which means switching off a single agent does not always end the incident. “Maybe you cut off one and now another one’s acting," Immler said.
Finally, Immler wants IT teams to resist the urge to clean up the agent’s environment before the investigation is complete.
“Don’t delete the agent immediately. Don’t wipe its environment. Don’t revert its settings. Don’t do things that can erase evidence that shows you how an attack worked,” he said.
“The overall goal should be to suspend, isolate, and delete only after you capture what you need.”