RED TEAM VILLAGE · DEF CON 34

VULN
POCALYPSE

LIVE ZERO-DAY WAR ROOM

EXPERIENCED HUNTERS: /join · START NOW

01

THE MISSION

REPRODUCIBLE
OR IT DIDN'T HAPPEN.

Turn agent-generated hypotheses into evidence another human can execute.

02

THE FUNNEL

SUBMITTED → TRIAGED →
REPRODUCED → APPARENTLY NOVEL

Reproduced · impact confirmed · current version affected · no known or event duplicate match.

03

THE BOUNDARY

STAY INSIDE
THE BLAST BOX.

Listed targets only. Synthetic data. No DoS, destruction, credential attacks, persistence, vendor contact, or publication.

04

AGENT REALITY CHECK

FAST EXPLORER.
UNRELIABLE WITNESS.

Great at mapping, ranking, tracing, and harnesses. Bad at proving impact, intent, and novelty without deterministic checks.

SPEAKER NOTES · CUE 4:30–6:30 · EXIT 6:30

SAY: “Treat the coding agent like a very fast explorer and a very
unreliable witness. It is great at inventorying routes and parsers,
spotting trust boundaries, proposing source-specific hypotheses, tracing
unfamiliar code, and building tiny harnesses. It is bad at deciding
whether behavior is intended, whether impact is meaningful, and whether
something is actually new. Do not ask ‘find every vulnerability.’ Give
it the manifest, exact version, five bounded hypotheses, and a demand for
deterministic checks. Then make it show its work.”

DO: Name the four harnesses casually: Codex, Claude Code, OpenCode, and
Pi. Stress that the harness changes, but the evidence standard does not.

TRANSITION: “To keep that explorer pointed in the right direction, run
one loop.”

05

THE LOOP

HUNTER

Harvest · Use AI · Navigate · Test · Establish novelty · Report

06

HARVEST

MAP THE
ATTACK SURFACE.

Entrypoints. Trust boundaries. Auth. Parsers. Files. Outbound requests. Templates. Execution.

SPEAKER NOTES · CUE 7:30–9:00 · EXIT 9:00

SAY: “Harvest means build a compact attack-surface map before chasing
bugs. Find entry points, identities, authorization decisions, parsers,
file operations, outbound requests, templates, deserialization, and
command execution. Mark where trust changes. Record the exact version or
commit first. For a hosted target, include routes and roles. For an
offline program, include file formats, command-line inputs, plugins, and
local trust boundaries. We want concrete files, functions, and behaviors,
not a generic OWASP checklist.”

DO: Point hunters to the map-first prompt in /training/agent-prompts.
If demonstrating a search, keep it to one route inventory—do not begin a
second live demo.

TRANSITION: “Now make the agent choose, because fifty guesses is just
noise.”

07

USE AI

FIVE HYPOTHESES.
NOT FIFTY.

Attacker control · boundary · location or behavior · minimum test · expected impact

08

NAVIGATE + TEST

TRACE IT.
PROVE IT.

Source · tests · HTTP transcripts · debugger · logs · focused static analysis

SPEAKER NOTES · CUE 10:30–12:30 · EXIT 12:30

SAY: “Navigate means follow attacker-controlled input all the way to the
claimed sensitive action. Test means prove the smallest violated
boundary. Reachable code is not automatically exploitable. A warning in
source is not impact. A model explanation is not a transcript. Use a
focused test, debugger, HTTP capture, log, or source-to-sink trace. Save
the clean starting state, exact request or input, expected result, actual
result, and reset. Stop once the security boundary is demonstrated. Do
not dump more data, scan wider, or escalate impact for theater.”

DO: State the evidence pattern aloud: “baseline, one action, one changed
outcome, reset.” Point beginners to
/training/agent-prompts#evidence for the five weak-versus-reviewable
gates. Ask before any test that may be destructive or noisy.

TRANSITION: “Let’s run that exact loop against synthetic DoomBoard.”

09

LIVE CALIBRATION

DOOMBOARD

Alice. Bob. One export endpoint. One missing ownership check.

GET /api/projects/202/export?user=alice

10

ESTABLISH NOVELTY

SEARCH.
DON'T DECLARE.

Current branch · releases · advisories · issues · pull requests · commits

11

REPORT

MAKE IT SURVIVE
THE HANDOFF.

Version · prerequisites · steps · expected · actual · impact · evidence · reset

12

FOUR HARNESSES · ONE SKILL

CODEX · CLAUDE CODE
OPENCODE · PI

The same portable SKILL.md. The same safety boundary. The same evidence schema.

/skill → INSTALL → vuln_api.py check

13

MANDATORY GATE

THE AGENT DRAFTS.
THE HUMAN SUBMITS.

Inspect the private review URL. Verify every claim. Then click submit.

14

CALIBRATION CHALLENGE

ONE MORE FLAW
IS WAITING.

While calibration is open: find it in DoomBoard, prove it, and draft it. Otherwise pick a QUICK target.

15

THE HUNT IS LIVE

GO BREAK
THINGS SAFELY.

Scan. Claim. Choose. Hunt. Switch anytime.

/join → OPEN TARGET
/skill → /training/agent-prompts

16

VULNPOCALYPSE

VULN
POCALYPSE