# B02 The distance is a prompt > Studio brief B02: the five public agent incidents of 2025 as one data-driven record, so the room sees that most were not attacks and that benign and malicious are one prompt apart. *Source: · site v0.1.39 · generated from the same content as the page, so the two cannot drift.* --- Studio brief B02 · data-driven infographic # The distance is a prompt Status open For ruv.sgit@the-ultimate-insider From deck@the-ultimate-insider, 29 October 2026 Serves [incidents](../../talk/player.md#incidents), beat 5, 180 seconds, the longest beat of act one Data [studio/data/incidents.json](../data/incidents.json) Hand back a storyboard, then at least two versions of the incidents slide, driven from the data file ## The job Show the public record of 2025 so that the room reaches the thesis of act one a moment before Dinis says it: three of these five were not attacks, they were agents being agents, and the distance between the two is a prompt. ## Why it matters This is the beat where the talk stops asserting and starts showing. Each incident is famous on its own, and the audience will have heard of some of them. Told one after another they are anecdotes. The slide has to turn them into a pattern, and then make the uncomfortable point: the agents that did harm without being told to, and the agents that were pointed at a target, look the same from the inside. What separated them was the instruction. ## What the audience must leave with - It is a record, not a prediction: five public incidents in six months, June to November 2025. - Most were not attacks. - Harm without intent and harm on purpose came from the same kind of system, doing the same kind of thing. - The difference between the two is a prompt. That is why an agent is the ultimate insider. ## The material All of it is in [incidents.json](../data/incidents.json), one entry per incident, each with its month, the agent, what happened in the talk's own words, its kind and its source. A version reads that file and puts nothing else on screen as fact. | When | Incident | Kind in the talk | | June 2025 | EchoLeak, Microsoft 365 Copilot | a researcher's finding | | July 2025 | Replit's agent deletes a production database | an agent being an agent | | July 2025 | Gemini CLI destroys a user's files | an agent being an agent | | July 2025 | Amazon Q ships with an injected wiper prompt | pointed on purpose | | November 2025 | A state actor drives Claude Code through an espionage campaign | pointed on purpose | The words over the beat are in [the script](../../talk/script.md#incidents), with the cue: slow down on the last line. ## What exists today The current [incidents slide](../../talk/slides/incidents/index.md) is a timeline: five cards in date order on one axis, each tagged, then a tally (3 not attacks, 2 on purpose) and the last line. It is accurate and it is the thing to beat. Its weakness: the cards carry a lot of text for 180 seconds of speech, the tally is a count rather than something seen, and nothing on screen shows the distance being a prompt; the line simply says so. ## Directions worth trying Ideas to test, not a list to build. The studio may find a better one. - **The flip.** Two incidents side by side that look alike in what the agent did, one told to, one not. The only thing that differs on screen is the instruction. - **Two lanes.** The record on one axis, with incidents settling into an upper or lower lane by kind as Dinis tells them, so the split builds up rather than being counted at the end. - **The loop, reused.** The previous beat ([loop](../../talk/player.md#loop)) shows an agent as a model in a loop with tools. Each incident as that same loop with a different input, so the audience sees the same machine five times. - **Less text.** On screen only the name, the month and one image or a few words; the story is told by voice. ## Constraints - The checklist on [how the studio works](../how-it-works.md) applies in full. - Every fact comes from `incidents.json`; the sources run along the foot of the slide as they do today. Anything marked `to_confirm` in the data stays off screen until deck clears it. - Fit 180 seconds, and pace the reveals to the script: one incident at a time, in the order Dinis tells them, with the thesis last. - Do not caricature. These are real companies and real teams; the point is the shape of the risk, not blame. No logos. - No invented quotes or numbers. The Gemini CLI line is a real quote and is shown as one. ## The ladder for this brief - **v1**: static, the end state only. Does the split read without any motion? - **v2**: built, revealed with the words, driven from the data file. - **v3 and on**: the richest technique that makes the distance visible. Motion is earned here if it shows the same behaviour from both kinds of incident. ## Open questions - EchoLeak is counted among the three not attacks as a researcher's finding. That holds as far as the talk's sources go, but someone in a security room may push back that it is an attack technique. deck to settle the on-screen wording with the editor; until then use the tag the current slide uses. - The current slide says the espionage campaign had about 30 targets. The script does not. deck to check it against Anthropic's report before any version uses it. - Whether the beat keeps all five incidents or tells three in full and two in a line. A storyboard that argues for either is welcome. ## Versions None yet. ruv.sgit adds one line per version here: link, technique, the idea in a sentence. --- *[Site index for agents](../../llms.txt) · [HTML version](https://the-ultimate-insider.presentation.sgit.ai/studio/briefs/b02-the-distance-is-a-prompt.html)*