Chapter 8: Incident Response

Outline

8.1 Incident Response for Agent Systems

• Why OpenClaw incidents are often authority, disclosure, and control-path failures rather than only host compromise
• Relationship between incident response, truthful operations, and bounded recovery
• The lifecycle this chapter defends: prepare, detect, assess, contain, investigate, eradicate, recover, review

8.2 Preparing Before the Incident

• Named incident roles, escalation authority, and out-of-band communication
• Preserving evidence sources in advance: logs, transcripts, tool events, route bindings, queue state, and deployment manifests
• Tabletop exercises tied to realistic OpenClaw failure modes

8.3 Detection, Assessment, and Severity

• Signals from audits, watchdogs, routing anomalies, duplicate ownership, provider abuse, and user reports
• Distinguishing operational defects from security incidents without dismissing either
• Severity based on authority at risk, data exposure, and ongoing attacker influence

8.4 Containment Without Destroying Evidence

• Isolating the right boundary: ingress path, worker pool, credentials, provider egress, or host
• Preserving chain of events while limiting blast radius
• Why panic shutdowns often make recovery less trustworthy

8.5 Investigation and Forensic Reconstruction

• Building a timeline from normalized events, transcripts, tool invocations, health signals, and deployment history
• Proving what happened rather than trusting model summaries
• Special handling for prompt injection, confused deputy, and duplicate-ingress incidents

8.6 Eradication and Recovery

• Replacing suspect runtimes, rotating secrets, re-establishing ownership, validating policy state
• Recovery gates before reconnecting production flows
• Monitoring for relapse after restoration

8.7 Human Escalation, Communications, and External Reporting

• When to escalate to human review immediately
• Internal coordination, legal/privacy involvement, and disciplined external communication
• Keeping communication truthful under uncertainty

8.8 Post-Incident Correction and Learning

• From lessons learned to concrete control updates
• Regression tests, policy revisions, and playbook repair
• Measuring whether the organization actually got safer

8.9 Case Study: Prompt Injection Becomes an Outbound Disclosure Attempt

• Step-by-step OpenClaw response walkthrough
• What containment and investigation look like in practice
• Where architecture shortens the incident

8.10 Chapter Summary

• The most important response principles
• Common mistakes to avoid
• Hand-off to resilience and future-hardening themes