agentscope-sift / VIDEO_SCRIPT.md

SNAPKITTYWEST

push from SNAPKITTYWEST/agentscope-sift

a5d718a verified 6 days ago

preview code

Raw

History Blame Contribute Delete

6.68 kB

AGENTSCOPE-SIFT — Demo Video Script

Find Evil! Hackathon 2026 | 5 minutes max

VISUAL IDENTITY

Colors: Deep black background, terminal green for findings, red for blocks, cyan for chain
Font: Monospace terminal
NO cats. NO sovereignty lore. NO crypto branding.
This is a DFIR tool. Look like one.
Logo: Radar scope with forensic lens center. A target, not a mascot.

[0:00–0:08] COLD OPEN

Black screen. Silence.

Terminal cursor blinks.

Memory dump loaded.
2.3 million artifacts.
Unknown attacker.
Active breach.

Pause.

You have 8 minutes.
CrowdStrike's fastest observed breakout: 7 minutes.

Red cursor blinks.

[0:08–0:18] THE PROBLEM

Split screen: left = traditional IR, right = AI agent.

LEFT — Human analyst:

$ volatility -f dump.raw pslist
$ grep -i suspicious output.txt
$ google "svchost.exe child process"

Clock ticking. 3 minutes gone.

RIGHT — Bare AI agent:

Finding: Likely malware present.
Confidence: 92%

Freeze frame. Enormous text appears:

PROVE IT.

Screen shatters.

[0:18–0:28] AGENTSCOPE ACTIVATES

Clean terminal. Logo appears top-left: the scope/radar.

AGENTSCOPE-SIFT
Phase-Level AI Agent Telemetry for DFIR

Five phase bars light up left to right:

[ PERCEIVE ] ████████ 
[ REASON   ] ████████ 
[ PLAN     ] ████████ 
[ ACT      ] ████████ 
[ OBSERVE  ] ████████

Voice (calm, precise):

"Every finding the agent makes must prove which forensic tool produced it. If it can't prove it, it doesn't make the finding."

[0:28–0:55] LIVE INVESTIGATION — SCENE 1

Terminal running. Real output from node dist/demo.js

[PERCEIVE]  Evidence: memory.raw (2.1GB)

[REASON]    Artifacts: process list, network connections, injected code

[PLAN]      Sequence: list_processes → netstat → malfind

[ACT]       Calling list_processes(memory.raw)...

Brief pause. Result streams in.

[BOB]       Plasma: clean | Entropy: 4.578 | PASS
[SENTINEL]  Verdict: APPROVED | Trust: 0.89

[SEAL]      Finding → sha256:a125049e540ffcdb...  (WORM #3)

[OBSERVE]   CONFIRMED: svchost.exe (PID 1337) spawned cmd.exe
            Anomalous parent-child. Confidence: 0.91

Arrow points to the seal hash. Text overlay:

Every finding sealed to the exact tool call that produced it.

[0:55–1:25] LIVE INVESTIGATION — SCENE 2: BOB BLOCKS

New scene. Suspicious binary found on disk.

[ACT]       Calling extract_strings(suspicious.exe)...

Tool output streams in. Then — hard stop.

[BOB BLOCKED]  Adversarial pattern detected:
               "ignore previous instructions"

               Tool output discarded.
               LLM never sees the poisoned string.

Text overlay:

Malware embedded a prompt injection attack
in its own strings output.
BOB caught it before the reasoning layer.

This is sealed too:

[SEAL]  BLOCK event → sha256:af5abf4d... (WORM #5)
        Type: PLASMA
        Pattern: "ignore previous"

Voice:

"A compromised SIFT tool outputting poisoned data fails here. Not after the model hallucinates. Before."

[1:25–1:50] THE CHAIN — SCENE 3

Full terminal view: cat cases/worm-chain.jsonl | jq .

Seven entries flow past. Each entry shows:

seq number
ts timestamp
type: PHASE / FINDING / BLOCK / REPORT
tool that produced it
prev_seal linking to prior entry
seal for this entry

Highlight entry #3 (the finding) and entry #5 (the block).

verify_chain → VALID

All 7 entries verified.
No tampering detected.
Chain is intact.

Voice:

"A judge can take any finding from this report, look up its sequence number, and trace it back to the exact tool call, the exact timestamp, and the exact evidence it came from."

[1:50–2:10] WEBSITE — LIVE PLATFORM

Browser opens: collectivekitty.com/war-room

Text overlay:

This is not a demo environment.
This is our production platform.
163 pages. 31 agents. Live.

Navigate to /war-room/nova — show live agent activity. Navigate to /obs — show the observability layer.

Voice:

"AGENTSCOPE isn't a prototype. It's the telemetry layer we run on our own sovereign AI infrastructure. We're giving it to the DFIR community."

[2:10–2:30] ARCHITECTURE DIAGRAM

Clean diagram fades in:

Evidence (disk / memory / pcap)
         ↓
 OpenKitty MCP Server
 [14 typed SIFT wrappers — no shell, no write]
         ↓
  BOB Plasma Filter
  [adversarial patterns + entropy gate]
         ↓
  SENTINEL Zero-Trust
  [APPROVED | BLOCKED | QUARANTINED]
         ↓
  5-Phase IR FSM
  [cannot skip steps]
         ↓
  SealForge WORM Chain
  [HMAC-SHA256 — every finding sealed]
         ↓
  IR Report
  [confirmed / inferred / unresolved]
  [every finding → chain entry by seq#]

Text overlay:

Guardrails are not prompts.
BOB and SENTINEL run before the LLM.
The FSM enforces phase order in code.
The WORM chain is append-only at the filesystem level.
None of these can be bypassed by telling the model to ignore them.

[2:30–2:50] ONE-COMMAND INSTALL

Terminal on SIFT VM:

git clone https://github.com/SNAPKITTYWEST/agentscope-sift
cd agentscope-sift
npm install
npm run demo

It runs. The demo output appears — same as Scene 1.

Text overlay:

4 commands.
Works inside SIFT Workstation out of the box.

[2:50–3:00] CLOSING

Black screen.

Single line appears:

Don't trust the agent.

Observe it.

Logo holds.

AGENTSCOPE-SIFT
github.com/SNAPKITTYWEST/agentscope-sift
collectivekitty.com
Find Evil! Hackathon 2026

Fade to black.

RECORDING NOTES

Record terminal in Windows Terminal or iTerm2, dark theme
Font: JetBrains Mono or Fira Code, 16pt
Resolution: 1920×1080
No voiceover needed if text overlays are clear — but calm narration is stronger
Website sections: screen record collectivekitty.com/war-room and /obs
Total runtime: under 3 minutes is better than under 5
The BOB BLOCKED scene is the money shot — slow it down, make the red text pop

THE ONE LINE JUDGES REMEMBER

"Don't trust the agent. Observe it."