Version: 2026-06-05
Summary
- On suspicion: stop, preserve state, report, do not keep experimenting.
- The first contact is the technical lead or maintainer.
- The developer may immediately stop the tool or agent and preserve state.
- No cleanup, no force push, no further experiments without the lead.
- Rollback must not be invented during the incident.
- If a secret was in AI context, assume compromise until properly assessed.
- Every relevant suspicion gets a short incident note.
- Incidents are classified simply: low, medium, high.
Topic
Handling suspected incidents involving AI tools, agents, third-party updates, rule packs, plugins, MCP integrations, Git, deployment, or external systems.
Starting Point
AI-related incidents are not always dramatic. They can start as a wrong command, accidental context exposure, generated credential, suspicious dependency proposal, unintended external write, or unexpected deployment behavior.
The important point is not to improvise while evidence disappears.
Risk
A developer may try to fix the situation quickly and destroy the information needed to understand the impact. Force pushes, cleanup commits, deleted logs, or repeated experiments can make the incident worse.
Recommendation
Stop the tool or agent. Preserve the current state. Inform the technical lead or maintainer. Do not continue experimenting until responsibility is clear.
The incident note should capture what happened, which tool was involved, what data or systems may be affected, which commands ran, which files changed, and which immediate containment step was taken.
Rollback paths should exist before incidents happen. Code rollback, database rollback, config rollback, and credential rotation are separate topics.
Review Checklist
- Was the agent or tool stopped?
- Was state preserved before cleanup?
- Were commands, diffs, logs, and prompts captured where appropriate?
- Are secrets treated as compromised if exposed?
- Is technical containment done before broad communication?
- Is there a rollback plan?
Wall rules
Stop, secure, report, do not continue experimenting.
Do not clean up before preserving state.
Rollback must not be invented during the incident.
A secret in AI context means: assume compromise first.
Contain technically first, then escalate wider.
Open questions for a project
- Who is the first incident contact?
- Where are incident notes stored?
- Which events are low, medium, or high severity?
- Which rollback paths are already documented?