By: E.J. Yerzak, Managing Director
For years, incident response plans have been built around a familiar assumption. When a cyberattack occurs, incident response teams will have time to investigate what happened, assess the evidence, and decide how to respond.
That assumption is becoming increasingly flawed.
Artificial intelligence is accelerating the speed and adaptability of cyberattacks. While the basic structure of an incident response program – clear reporting lines, defined responsibilities, escalation procedures, legal and compliance involvement, and coordinated communications – remains essential, the operational model behind many response plans is becoming outdated.
The problem is simple. Cyberattacks are occurring faster than humans can investigate them.
From Human Speed to Machine Speed
Traditional cyberattacks often unfold over hours or days. An attacker might repeatedly attempt credentials, gain access to an email account, create forwarding rules, move laterally through a network, or establish persistence. Security teams generally had some time to react.
Alerts could be reviewed. Logs could be analyzed. False positives could be ruled out. Incident response teams could convene, assess the available evidence, and determine the appropriate next steps.
AI-enabled attacks can compress that timeline dramatically. An AI agent does not necessarily need to wait for a human attacker to interpret an error message or determine what to try next. It can test an approach, evaluate the result, change tactics, and immediately pursue another path.
More importantly, AI agents can be goal-oriented rather than process-oriented. A traditional automated attack may follow a relatively static sequence of instructions. If one step fails, the process may stop. An AI agent can potentially react to that failure, select an alternative technique, and continue toward the same objective. The result is an attack that branches dynamically instead of progressing through a predictable sequence.
By the time a security analyst understands what happened several minutes ago, the attack may already have moved elsewhere.
AI Is Expanding the Attack Surface . . . to Other AI
The rapid adoption of enterprise AI tools creates another issue we should have seen coming. The AI environment itself increasingly contains sensitive information.
Employees are using AI platforms to summarize contracts and side letters, analyze financial information, draft communications, review portfolio models, process internal documents, and perform other business functions. As more proprietary and confidential information enters these systems, they become increasingly attractive targets.
Traditional data loss prevention controls may also struggle with AI-generated output. Many DLP systems are designed to identify known patterns, such as Social Security numbers, account numbers, or other structured data. AI systems, however, can express the same information in different forms. A Social Security number that would normally appear as a numeric string could instead be expressed in words or embedded within natural-language output. The sensitive information is the same, but its format has changed. That creates a challenge for controls designed primarily around static pattern matching.
The Threat Can Adapt
Perhaps the most significant change we are seeing is that threat actors are demonstrating an ability to modify their behavior in response to defensive actions. An AI-enabled attacker that encounters an access restriction may attempt a different authentication method. Malware code may be rewritten on the fly to evade newly implemented detection rules. Multiple agents may test different attack paths simultaneously.
Security teams may no longer be responding to a fixed threat. They may be responding to a threat that is actively adapting on multiple fronts to the response itself.
That malicious use of AI requires a different incident response mindset to combat it.
Four Ways Incident Response Needs to Change
- First, containment needs to become more dynamic. Organizations cannot rely solely on static indicators of compromise. Detection and containment increasingly need to focus on behavior and adapt as attack techniques change.
- Second, firms should consider whether certain high-confidence indicators warrant immediate defensive action. Historically, an alert might be sent to an analyst for investigation before credentials are disabled or network access is restricted. In the age of lightning- speed attacks, that delay may be too long.
Organizations should evaluate when systems should automatically revoke credentials, invalidate authentication tokens, isolate endpoints, or block malicious traffic while an investigation continues. In other words, we may need to use AI on the defense to have a chance at stopping AI used by the offense. The objective is not to automate every response. It is to identify situations where the risk of waiting exceeds the risk of temporarily disrupting legitimate activity.
- Third, deepfakes and AI-enabled impersonation should be incorporated into incident response planning. During a crisis, attackers may attempt to impersonate executives, employees, vendors, or members of the response team using convincing voice, video, or written communications.
Organizations should establish alternate communication channels and out-of-band identity verification procedures before an incident occurs. A familiar voice or convincing video should no longer be treated as sufficient authentication.
- Finally, response procedures should account for the limitations of simply resetting passwords.
If an attacker already possesses valid authentication tokens or active sessions, changing a password may not terminate access. Depending on the circumstances, firms may need to terminate active sessions, invalidate tokens, reset credentials, and require multifactor authentication methods to be re-registered.
War-Game the Attack and Change Your Assumptions
Organizations should also reconsider how they conduct cybersecurity tabletop exercises. Instead of asking only, “What would an attacker do next?” response teams should ask what an AI agent might try next. And how many approaches could it pursue in parallel at the same time?
Tabletop exercises should include attacks that branch unpredictably, adapt to defensive actions, manipulate communications, and move faster than the response team can comfortably deliberate. This doesn’t mean tabletop exercises are dead. Quite the opposite. It means tabletop exercises take on an even greater significance to condition thinking critically and creatively about responses.
The timeline for attacks should also be compressed. If a tabletop exercise gives the team an hour to evaluate every development, it may be testing a response model that does not reflect the emerging threat environment.
- What happens if the team has five minutes to decide? What happens if the team has five seconds?
- What happens if the attacker changes tactics before the team reaches consensus?
- What happens if several seemingly unrelated alerts are actually coordinated attempts to achieve the same objective?
Those are increasingly the questions incident response teams need to be prepared to answer.
The Playbook Still Matters. It Just Needs an Upgrade.
Like tabletop exercises, incident response plans are not obsolete.
But many were designed for an environment in which attackers and defenders operated at roughly the same speed. It’s a classic cat and mouse game between hackers and incident response teams. AI is changing that equation and has given the mouse a jet pack. Incident response teams need a similar boost.
The next generation of incident response will require the same strong governance organizations rely on today – an interplay between compliance, operations, legal, and cybersecurity teams. These governance and risk management structures must be combined with faster detection (including defensive AI), adaptive controls, carefully designed automation, stronger identity verification to mitigate deepfakes, and the ability to contain a threat before every unanswered question has been resolved.
We don’t just need to respond faster. We need to build an incident response capability that can operate at the speed of the threat.
With any luck, by the time we update our incident response plans, it will be time to do it all over again.