Crisis & riskAgenticAugust 8, 2026

The Deepfake Dry Run

Suggested tool: Claude (best at refusing to assume capabilities you did not give it, which is the entire point of this exercise)

The scenario

A synthetic video of your CEO is circulating. It is good. A reporter has already emailed. You have 60 minutes today to build the muscle you will need on the day you actually need it.

The prompt

Act as a crisis communications advisor running a tabletop exercise. Below is a scenario in which a convincing synthetic video of our senior executive is circulating, along with an honest inventory of our current verification capability and approval chain.

Produce five things. First, a 90-minute response timeline broken into 15-minute blocks, starting from the moment we become aware. Second, a verification chain with a named role responsible at each step and a realistic time estimate for each step. Third, a holding statement we could send in the first 30 minutes that does not confirm or deny anything we cannot yet verify. Fourth, a platform escalation and takedown sequence in priority order. Fifth, the three questions a reporter will ask that we currently cannot answer.

Work only from the capabilities I actually listed. Do not assume we have forensic tools, a standing legal contact, or 24-hour monitoring unless I said so. Where a step depends on something we do not have, say so plainly and tell me what the workaround costs us in minutes.

How to use it

Pick your most exposed principal. Usually the CEO, sometimes the founder, occasionally the scientist who does all the podcasts. Write down three things: how much public video and audio of them exists, who inside your company could technically confirm or deny a clip, and who has authority to approve a statement at 11pm on a Saturday.

Write one paragraph describing the fake. Be specific. What does the clip show, where did it surface first, and who is amplifying it. Vague scenarios produce vague plans.

Paste both into Claude or ChatGPT with the prompt below.

Read the verification chain with a stopwatch mentality. Add up the realistic time estimates. If the total is longer than the time it takes a story to publish, that number is your actual finding, not the plan.

Save the holding statement somewhere your on-call person can reach it without a laptop. Then put a 20-minute re-run on the calendar for next quarter with a different scenario.

Why it works

It turns an abstract fear into a timeline with names on it, which is the only version of a crisis plan that survives contact with a real crisis.
It exposes your verification gap in an hour instead of during the incident, and that gap is almost always the real story.
It gives you a holding statement written by a calm person, because the one written at 11pm by a panicked person is the one that gets screenshotted.

Cowork version

For Claude Cowork, where the prompt can read the files in a folder.

Run a deepfake tabletop exercise using the files in this folder.

1. Read the crisis plan and the contact list. Build an honest inventory of what we can verify today, who can authorize a statement, and how long each of those steps realistically takes after business hours.

2. Read the public footprint material and assess how much source audio and video exists for our principal. Note which formats would be easiest to convincingly fake.

3. Write three distinct scenarios at increasing severity. A fake clip on a fringe platform, a fake clip picked up by a trade outlet, and a fake clip a national reporter is calling about.

4. For each scenario, produce a 90-minute timeline in 15-minute blocks, a verification chain with named owners from our actual contact list, and a holding statement.

5. Flag every point where the plan depends on a capability we do not have in the source files. Do not invent tools, vendors, or roles that are not documented. Rank those gaps by how much time each one costs us.

6. Save the three scenarios as one tabletop packet, the holding statements as a separate one-page card, and the gap list as a short brief I can take to leadership.

This prompt first ran in TIC. Read the original issue →