# Containment Under Uncertainty > Containment Under Uncertainty - Skill: `annatchijova/containment-under-uncertainty` (Agent Skill) - Install (CLI): `npx skillmds@latest add annatchijova/containment-under-uncertainty` - Raw SKILL.md: https://api.skillmd.com/api/skills/annatchijova/containment-under-uncertainty/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: annatchijova (https://skillmd.com/u/annatchijova) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/annatchijova/containment-under-uncertainty --- # Containment Under Uncertainty The defining condition of containment is that you must act before you know the scope, and both acting and waiting cost you something. Evict too early and loudly and you warn the adversary — who then burns their access down, destroys evidence, or accelerates to finish the job before you finish yours. Wait for certainty and the data keeps leaving. There is no move that is free of one of these costs, so the discipline is not "decide once you understand it"; it is **buy information without buying disaster**, and reason about every action as reversible or irreversible. The two failure modes are symmetric, and organizations reliably pick one by temperament: - **Premature loud eviction.** Pulling the plug, mass-resetting, wiping the box — before you know what it connected to. You destroy the forensic scope you needed (`what else did this credential touch?` is now unanswerable), and you tell the intruder you are onto them while they still hold three other footholds you never found. - **Analysis paralysis.** "We don't fully understand it yet" hardens into inaction while exfiltration continues. Perfect scope is never available during the bleed; the demand for it is how a team does nothing and calls it caution. Composes with the library: - **assume-breach-modeling** — you contain against the *assumed* blast radius, not the confirmed one; this supplies it - **irreversible-action-gate** — wipe, rebuild, rotate-everything, power-off are one-way doors; they get the gate, not a reflex - **incident-timeline-reconstruction** — containment must preserve the evidence this later depends on; do not trade the timeline for a fast eviction - **credential-material-triage** — deciding what to rotate first is a blast-radius question about the captured material - **honest-degradation** — the decision record states the scope you had vs assumed; a WARN is not a PASS - **dual-use-behavior-adjudication** — before you contain, be sure the "intrusion" is not your own admin (containing a false positive has real cost too) --- ## Step 1 — Sort every option into reversible and irreversible Before choosing, classify the moves, because the classification is most of the decision: - **Reversible / low-regret** — network-isolate a host (you can un-isolate it), block an egress domain, disable one account, increase logging, snapshot memory. These buy safety and information and can be undone if you were wrong. - **Irreversible / high-regret** — wipe or rebuild a host (evidence gone), power it off (volatile state gone — see `acquisition-order-of-volatility`), rotate every credential at once (you may lock out the responders and tip the adversary simultaneously), publicly disclose. Default to the reversible moves *first*: they narrow the bleed and preserve your options. The irreversible ones require the gate in Step 4. --- ## Step 2 — Price the observation-versus-eviction tradeoff explicitly Watching a live intrusion has value (you learn scope, TTPs, other footholds) and a cost (data leaves while you watch). Evicting has value (bleed stops) and a cost (you lose the intel and warn the adversary). State which side the specific incident favors: - **Lean toward observe** when the ongoing loss is bounded and slow, the adversary seems unaware, and you have not yet found the full foothold set — the intel prevents a whack-a-mole where you evict one access and miss two. - **Lean toward evict now** when active, fast, high-impact loss is underway (ransomware staging, bulk exfil in progress, hands-on-keyboard destruction) — intel is worthless if the patient dies during the diagnosis. This is a judgment with an owner and a rationale, recorded — not a reflex, and not a decision made by whoever is loudest in the bridge call. --- ## Step 3 — Contain against the assumed blast radius, not the confirmed one You will never have confirmed scope in time. Use `assume-breach-modeling`: assume the compromised identity reached everything its privileges allow, and contain *that* surface, while marking clearly that it is assumed. This is how you avoid the whack-a-mole trap — evicting the one host you found while the reused credential still works on twenty others. Rotate/isolate by blast radius (via `credential-material-triage`), highest-reach first, in an order that does not lock out your own responders mid-eviction. --- ## Step 4 — Put the one-way doors behind the gate Wipe, rebuild, mass-rotation, and power-off are irreversible; they must not be reflexes under adrenaline. For each, before pulling it: - Confirm the evidence you need has been acquired first (memory before power-off; image before wipe — `acquisition-order-of-volatility`). - Confirm it is coordinated: a partial mass-rotation that misses one system hands the adversary a still-valid key and announces the response. Eviction works best *simultaneous and complete*, which requires knowing the foothold set — the very thing Step 2's "observe" side was buying. - Record the decision, its rationale, and the accepted residual uncertainty (`irreversible-action-gate`). --- ## Step 5 — Preserve the scene while you contain, and write down what you didn't know Containment and forensics are in tension but not opposed: isolate rather than wipe, snapshot before you change, log the footprint your own actions leave. The incident timeline (`incident-timeline-reconstruction`) is built from evidence a hasty eviction destroys — so the honest close of every containment action is a note of the scope you actually had versus assumed, and what a later analyst has lost because you had to act before you knew. That WARN is not an admission of failure; it is what makes the next decision in the same incident an informed one. --- ## The one-line test If your containment plan requires knowing the full scope first, you have designed a plan for an incident you are no longer in. Take the reversible moves now against the *assumed* blast radius, keep the one-way doors behind the gate until the evidence is saved and the eviction can be complete, and write down — as you go — the difference between what you knew and what you assumed.