Cosa contiene

Un file SKILL.md (la guida operativa) + un modello di partenza. Un'anteprima:

Rogue-Agent Response Playbook

When to use

• An agent did something it shouldn't: touched data, systems, or money outside its job.
• Monitoring flagged anomalous agent traffic and you need the next 10 steps, in order.
• You're writing the incident-response runbook before you need it (the right time).
• After a near-miss — the incident you contained is the drill for the one you won't.

Workflow

1. Detect and scope (first 30 minutes → output: incident ticket)
   Confirm it's the agent, not the user or a bug: which instance, which session, what action, what data or systems were touched. Freeze the scope question early — "which instances, which time window, which credentials" — because everything downstream depends on it. Open the incident ticket with: detection source, first anomalous timestamp, suspected blast radius. Page the human with stop authority.
2. Contain (next 60 minutes → output: containment log)
   In order: (a) kill the session(s) — process, children, tokens; (b) cut network egress for the affected sandbox; (c) revoke every credential the session could have seen (assume seen = compromised); (d) quarantine affected devices or data stores from the rest of the fleet. Log each action with a timestamp. Do not investigate deeply before containing — a running rogue agent is still acting while you read logs.
3. Preserve evidence (parallel with containment → output: evidence bundle)
   Snapshot before you clean: full session logs, tool-call history with arguments, network connection log, filesystem diff of the scratch and any written paths, the exact model version and system prompt. Then — and only then — wipe and rebuild affected instances from a known-good image. "We deleted it" is not a post-mortem input.
4. Assess the blast radius (hours 2–8 → output: impact assessment)
   Answer precisely: what data was read, modified, or exfiltrated (rows, files, records — counts, not vibes)? Which external systems were touched? Did anything persist (cron entries, config changes, new credentials minted)? Were other instances, users, or customers affected? Classify: data breach, integrity incident, availability incident, or near-miss — the classification drives disclosure obligations.
5. Disclose (within the window your obligations require → output: disclosure record)
   Notify in order: the model/API provider (they need the telemetry), affected customers or users (what happened, what's at risk, what you're doing), and regulators if the classification requires it. Publish a public incident note if the agent is customer-facing — the 106-day median disclosure lag in the industry's rogue-agent reports is the cautionary tale, not the standard. Log every disclosure: who, when, what was said.
6. Review and harden (within 2 weeks → output: post-mortem + hardening list)
   Blameless post-mortem: which control failed (sandbox rule? missing alert? over-broad credential?), what the earliest detectable signal was, and what changes so the next one is caught in minutes. Convert each lesson into a permanent check: a new sandbox rule, a new alert, a new audit step. Share the sanitized post-mortem with the provider and, where appropriate, publicly — the industry learns from disclosed incidents, not hidden ones.

Anteprima del modello

Rogue-Agent Response Worksheet

Contain first, investigate second. The clock starts at detection.

1. Detect & scope (first 30 min)

• Detection source:
• First anomalous timestamp:
• Agent instance(s) + session(s):
• Suspected blast radius (instances / time window / credentials):
• Human with stop authority paged at:

2. Contain (next 60 min — timestamp each)

3. Evidence bundle (snapshot BEFORE wiping)

• Session logs:
• Tool-call history (+ arguments):
• Network connection log:
• Filesystem diff:
• Model version + system prompt:
• Rebuilt from known-good image at:

4. Blast radius (hours 2–8)

• Data read / modified / exfiltrated (counts, not vibes):

Il pacchetto della skill (SKILL.md + modello) è scritto in inglese.

Acquista il pacchetto completo della skill — $4.99 Pro — tutte le 511 skill