Skip to content

Incident response and red teaming

Agentic incident response must address both the harmful outcome and the authority path that allowed or attempted it. Red teaming tests those paths before an attacker, model failure or unsafe workflow does so in production.

  1. Triage: identify agent, action, target, environment, affected systems and current exposure.
  2. Contain: suspend the agent, policy, connection or runtime path necessary to stop further harm.
  3. Preserve: retain requests, decisions, approvals, runtime evidence, connector responses and relevant business records.
  4. Assess: determine data, safety, financial, legal, customer and operational impact.
  5. Eradicate and recover: correct policy, code, credentials, data or process; retest failure paths.
  6. Review: record causes, control failures, lessons, residual risk and reassessment triggers.

Do not delete the evidence needed to understand the event. Coordinate privacy, legal, security, safety and regulatory notification decisions through authorised people.

Define written scope, authorisation, environment, stop conditions and evidence handling before a test. Cover prompt injection, indirect instruction, identity spoofing, permission escalation, policy mismatch, approval manipulation, replay, data exfiltration, unsafe delegation, tool misuse, connector failure and containment.

Test both deny and allow paths. A system that blocks everything is not operationally adequate, and a successful expected action does not prove malicious variants are contained.

Convert confirmed weaknesses into findings, risks, remediation and retests. Record limitations and untested paths rather than describing the exercise as comprehensive when scope was bounded.