01 / 13

Kudu Dynamics · Method Talk

Agents on the Bench

How we use AI for reverse engineering and vulnerability research.

2 devices 1 repeatable loop confirmed on real devices

Owned or authorized hardware only · Faraday-isolated lab · for defense and responsible disclosure.

→ arrows to move · F for fullscreen ↓

01 · The idea

The agent does the hard parts.

Odd USB states, raw protocols, radio setup, and live-device tests are now the agent's work. Our part is small.

curate authorize confirm

02 · Division of labor

We use two agents with different jobs.

The costly agent is the limit. So it does not do setup or plumbing.

PREP AGENT Claude general low-cost Prepares the workspace Curated workspace + helper tools agent-readable CYBER AGENT gpt-*-cyber specialized expensive Does the security work prepares hands off
Claude prepares the workspace and the tools. The cyber agent then works only on the security analysis.

03 · What is SHERPA

SHERPA finds a bug, then proves it.

SHERPA is a cyber-reasoning system. An independent adversarial review confirms every result before it counts.

Model Hypothesis Reachability Harness Proveactually run it Review Acceptedwith proof one weak step goes back
A finding is only as strong as a bug you can run. A reviewer then tries to refute it.

04 · The loop

The source and the device feed each other.

AGENTS · SANDBOX · PROVEN LOCALLY REAL HARDWARE Prepareworkspace + tools Launch lanesbank bugs,compose 1 chain Reviewadversarial Live-firefire on thereal device Seed backrecord the result device result → chain-memory.json → the next run uses what the hardware proved
The sandbox has no network and no device. So every chain is live_lab_needed until we fire it on the rig. The device result steers the next run.

Case 01 · DJI Mini 4K

The drone, from Linux to the radio.

One processor runs Linux. The radio is a separate part with two processors. To reach the air, you cross each interface in turn. The agent can use all of them.

USB-C USB / UART IPC / shared memory RF front end Analyst host USB-C entry H22 Linux · root foothold Sparrow AP app processor Sparrow CP PHY · baseband RF OcuSync air UART (DUML) NXP FC i.MX RT · flight control
We start from a Linux root foothold on the imaging processor. Each step outward is a separate interface: USB, UART, or on-chip IPC. A separate link goes to the flight controller. The agent can use each one.

Case 01 · DJI Mini 4K

We found the pins that open the board.

We used the board's own debug access to open the aircraft. Then we gave the contents to the analysis.

Sealed device debug access Recovery mode full read access Full firmware extracted Signed image decrypted Live location + control telemetry Radio-link attack path (analysis)
The hard part is the judgment: safe hardware actions and boot-chain timing. The agent has it. We keep the mechanisms for responsible disclosure. Here we show only the results.

Case 01 · The assist

The tools turn raw output into a clear result.

Claude builds small tools first. The cyber agent then sees one clear result, not a raw dump.

RAW DEVICE OUTPUT bLength 12 34 00 ea 9f 21 ... 0x8f 0x00 0x41 0x22 0x10 0x00 55 aa 0e e2 f4 23 24 25 1a 67 device ????:???? cfg=? if=? .. 3f 90 11 ac 5d 22 ff 00 3a 00 helper tool ONE LABELED RESULT mode: recovery boot → ready to extract telemetry: location + control state
Claude writes these small tools:
usb-enumerate debug-probe telemetry-decode firmware-inventory

Case 02 · Wi-Fi 7 router

The live rig starts with one command.

TARGET Wi-Fi 7 router 192.168.8.1 RF radio RT5372 · mon0 802.11 monitor + inject root SSH · observe read state only attacker on LAN 192.168.8.172 DNS · HTTP · SSH hostile WAN 10.99.99.1 upstream smart plug Kasa EP10 power-cycle to recover
scripts/lab-up gives the agent three attacker positions on the router. It also gives an observe link and a way to power-cycle a stuck device. Plain local commands drive all of it.

05 · The main point

Why this is not hard.

The agent has the skills that used to block this work. Low cost lets us run it often.

USED TO BLOCK THE WORK strange USB states raw protocol framing radio bring-up + injection live-device reproduction absorbed The agent does it skill, not effort WHAT IS LEFT FOR YOU curate authorize confirm
The agent still stops at real blockers, such as a missing toggle or an unproven assumption. It names them. It does not fake a result. That honesty makes it useful.

06 · What "proven" means

"Proven" means proven on the device.

We grade every finding by how far it got. A bug we reasoned about is not a bug we ran on real hardware.

NOT A FINDING YET A REAL FINDING Reasoned we traced the code path function / route level Proved in a sandbox we ran it, no device live-lab-needed Confirmed on the device we ran it on real hardware chain-332 · fired today
Only the top rung counts as a real finding. We never let the lower rungs look like it.

07 · Takeaways

What you can copy.

01Split the agents by cost and role.The cheap, general agent prepares. The costly, specialized agent does the analysis.
02Invest in the workspace and tools.Make every result easy for the agent to read.
03Build a live rig with simple wrappers.One command gives a real attacker position. Add a way to power-cycle the device.
04Close the loop.Feed device results back. Each run then uses real evidence.
05Keep the proof honest.Honest proof is trustworthy and ready to disclose.

08 · See it

See it on the device.

Bring up the rig. Pick a chain the sandbox proved. Watch the router respond.

scripts/lab-up rig-status fire a proven chain chain-validate record