Skip to content
Beyond Prompt AI Studio
AI Security for Businesses

Practice: The Incident Response Roadmap

The previous modules of this course showed how to prevent risk. This module answers the question that eventually comes up anyway: what do you do when an AI agent actually does something unexpected? An established four-stage framework provides the structure – adapted to the signals AI systems specifically produce.

Four incident signals – worth remembering

Try it yourself: classic IT incident vs. AI-layer incident

Tap a dimension to compare both sides.

Classic IT incident

Classic IT incident: compromised credentials, unusual network traffic.

AI-layer incident

AI-layer incident: unusual tool-call patterns, prompt patterns.

The four-stage framework: not reinvented, just adapted

According to the US standards body NIST (framework SP 800-61, most recent revision from April 2025), every security incident runs through four phases: preparation, detection and analysis, containment/eradication/recovery, and post-incident activity. This framework is well established and doesn't need reinventing for AI incidents – it only needs to be supplemented with AI-specific signals a classic IT security team would otherwise miss.

Why AI-layer incidents produce different signals

A classic IT incident typically shows up through compromised credentials or unusual network traffic. An AI-layer incident – prompt injection in a production application, an agent's escalating tool calls, an attempt to extract a model through targeted queries, prompt-based data exfiltration – often looks completely unremarkable on classic monitoring dashboards, because it disguises itself as normal application use. Anyone only looking for classic signals reliably misses this incident class.

The four phases, applied to AI incidents

For AI systems, preparation means: logging what an agent actually does (not just what it answers), before an incident is even suspected. Detection and analysis means: specifically looking for AI signals – unusual tool-call patterns, repeated failed instruction attempts, unexpected data volumes in outputs. Containment for an agent often means: stopping the agent and immediately revoking its tool privileges, not just blocking a network account. Post-incident activity means: adjusting the underlying rights structure or input checking so the same incident type doesn't recur identically.

Practice section: the roadmap before the first incident, not after

The most important practical advice: this roadmap belongs on paper before the first incident happens, not while it's underway. Concretely, that means: decide in advance who gets informed in an emergency, how an agent can be technically stopped in an emergency (see the architectural defense from an earlier module in this course), and which logs need to exist beforehand at all to reconstruct what happened afterwards. A company that only answers these questions during an ongoing incident loses valuable time at exactly the moment it's most expensive.

The key points

  • The established four-stage NIST framework (preparation, detection/analysis, containment/eradication/recovery, post-incident activity) doesn't need reinventing for AI incidents, just adapting.
  • AI-layer incidents (prompt injection in production, tool-call escalation, model extraction, prompt-based data exfiltration) produce different signals than classic IT incidents and are often missed by classic monitoring.
  • For AI systems, preparation means logging what an agent actually does, not just what it answers.
  • Containment for an agent often means immediately revoking tool privileges, not just a network block.
  • The roadmap belongs on paper before the first incident – anyone who only clarifies these questions during an ongoing incident loses valuable time at exactly the moment it's most expensive.

The second OpenAI incident: for five days it was an attack. Then it was an accident.

Quick check: did it sink in?

1 / 3

Why do classic monitoring dashboards often miss AI-layer incidents?

Want to set up an incident response roadmap for your own AI systems?