BrighouseSec · product security for drones and autonomous systems

Command-link evidence an auditor will accept.

Whether your vehicle accepts unsigned commands, which links are exempt from signing, and what a second control plane exposes that a signing key never reaches. Measured on PX4 and ArduPilot, not asserted, and mapped to the Cyber Resilience Act.

Fixed price. Written authority before anything is touched. Raw records with every report.

Signing key activeArduPilot Copter 4.7.2 · as measured

Same aircraft. Same key. One port still answers anyone.

signed frameunsigned frameSee the measurement ↓

For drone and autonomous-systems manufacturers

Your radio on the wrong port undoes the signing you just turned on.

On ArduPilot, the exemption that lets an unsigned command through follows the channel number, not the kind of link. We measured it: a key-active vehicle answered an unsigned read on channel 0 over TCP and over UDP alike, and refused the same read on channel 1 — including when channel 1 was running MAVLink 1, which the documentation says is unprotected. Twenty-three cases, four refusals, none inconclusive. If your telemetry radio sits on channel 0, that is your exposure, and signing did not close it.

This is documented upstream, not something we found: ArduPilot issue #34489, with fix PR #34780 still unmerged as of November 2026. What we add is the independent measurement, a formal model that reproduces it, and the mapping to the regulation. The raw records ship with the report so your own engineers can re-run every case.

Secure-by-default score · measured, 0 to 100

Where does your command link sit?

No signing key. Every link answers anyone.

14, 42 and 66 are measured on ArduPilot 4.7.2 and PX4 1.18 in simulation; 82 is the metric's ceiling. How the score works ↗

Signing key Arrives on Dialect

Measured: 23 cases, 4 refusals, none inconclusive. The exemption follows the channel index, not the transport.

47accept and reject decisions recorded while the formal model was checked against both stacks. The model agrees with all forty-seven.Acceptance Semantics for MAVLink 2 Signing
23of those decisions carry the discriminating evidence, with seven refusals observed. Twenty are keyless and therefore tautological, and PX4 1.17 with no signing is the negative control.Same record, published beside the model

Measurement

MAVLink & C2-link audit

Whether your vehicle accepts unsigned commands, falls back silently to legacy MAVLink 1, or answers peers that presented no credential — the exact things the C2 class mark says the link must resist. Run against your own firmware configuration.

  • Unsigned-command and downgrade tests
  • Attack-surface and service exposure
  • Reproducible report an auditor can read

From GBP 1,200 · one week

Evidence

Drone CRA Technical File

The Annex I evidence pack for one product: your firmware and your parameters under test, every finding mapped to the requirement it answers, a CycloneDX firmware SBOM, a vulnerability-reporting runbook — and an explicit list of what was not covered.

  • Annex I Part I/II traceability
  • Secure-by-default score as headline KPI
  • Raw data and scripts, so you can repeat it

From GBP 1,800 for a single product (ten working days) · GBP 3,500 and up where there are several configurations or a companion computer (three to five weeks)

Verification

Re-test after you fix it

You change the configuration or carry the patch; the same measurements run again and the file carries the new numbers beside the old ones. Cheaper than the first pass, because by then your system is already understood.

  • Same harness, same cases
  • Before and after in one document
  • Scoped in advance, no surprises

From GBP 600

The part a signing key does not reach

There is a second way into the flight stack.

Signing protects MAVLink. PX4 also ships a uXRCE-DDS bridge that reaches uORB directly and is unauthenticated by design. On 4 November 2026 we measured it offering agent discovery to whichever process holds the port, on both v1.17.0 and v1.18.0-rc1. A key provisioned over MAVLink does nothing there.

One qualification, because it changes what the number means: those measurements are POSIX/SITL builds, which is not NuttX. The bridge is started by the POSIX startup script; the stock NuttX one has no reference to it. We say that in the report rather than letting you assume the worse case.

What is already in force

NOWCRA Article 14, since 11 September 2026. Actively exploited vulnerabilities reported to ENISA's single platform within 24 hours. Not a future date.
NOWDelegated Regulation 2019/945, Annex Part 3. The C2 link must be protected against unauthorised access. This one does not wait for the CRA.
2027Annex I essential requirements, from 11 December 2027.

The obligation sits with the manufacturer at the point of sale, not with whoever flies the aircraft afterwards. And a drone's C class does not determine its CRA class — they are different axes, and conflating them is the kind of mistake a legal adviser catches on the first call.

Who this applies to

183 manufacturers surveyed. 27 confirmed selling into the EU.

A class C2 or C3 mark already requires a command link that is protected against unauthorised access, and the Cyber Resilience Act adds evidence duties on top. This is our own survey of who builds these aircraft, November 2026.

Click a pin — or drag the globe. Each pin is a country, and its height is how many drone manufacturers we found there. Our own survey, November 2026.

makers found confirmed EU-selling pin height = how many

Published work

The method is in the open, including where it failed.

Four preprints, an open auditing tool with 53 tests, and the raw measurement records published beside them rather than summarised away. Every one of these is a preprint and has not been peer reviewed — it says so on the page, because a reader deserves to weigh it accordingly.

  1. 01
    Signed but Unprotected: MAVLink 2 message-signing defaults and their bearing on the EU Cyber Resilience Act ↗

    Every default link was unsigned, answered unauthenticated reads and exposed MAVLink FTP. With signing switched on, ArduPilot's primary channel signs every outbound frame while still accepting unsigned commands — so a purely passive observer would wrongly judge it protected.

    Preprint · not peer reviewed10.13140/RG.2.2.20818.31689

  2. 02
    A Decade of Defaults: how "secure by default" evolved in ArduPilot and PX4, 2016–2026 ↗

    Signing has never shipped on by default in either stack. The unexpected result: the binding constraint on auditing old releases is reproducibility, not posture — pre-4.5 ArduPilot no longer builds on a current toolchain, which bears directly on the CRA's expectation that products stay patchable.

    Preprint · not peer reviewed10.13140/RG.2.2.27660.27529

  3. 03
    Acceptance Semantics for MAVLink 2 Signing: bounded symbolic analysis and model-based conformance testing ↗

    A formal model of when a frame is accepted, checked against real builds: it agrees with all 47 recorded decisions. One pair of configurations is a controlled experiment, not a coincidence — moving the same transport between channel indices flips the decision exactly as the trusted-link parameter predicts.

    Preprint · not peer reviewed10.13140/RG.2.2.20949.38885

  4. 04
    Scoring "Secure by Default": a reproducible, CRA-grounded metric ↗

    Turns an endpoint's observable posture into a 0–100 score with a letter grade, on three axioms — provenance, monotonicity and CRA-grounded weights — released as code with a property-test suite that checks them. Grade B is the ceiling for MAVLink 2 signing alone.

    Preprint · not peer reviewed10.13140/RG.2.2.34371.16164

  5. →
    mavaudit — the auditor itself, with the raw records ↗

    Passive and key-less, read-only by construction. The models, harnesses and every measurement record ship with it, including the cases that came back inconclusive and earned nothing.

    Open sourcegithub.com/brigghouse/brighouse-autonomy-lab

A number worth having before you buy rather than after. A MAVLink link configured as well as signing alone allows scores 82 out of 100, a grade B — never higher. The last eighteen points are not in the protocol's gift: twelve need keys distributed out of band, because the key-exchange-over-the-link flow both stacks ship is the trust-on-first-use window, and six need an encrypted radio or a VPN underneath, because MAVLink 2 authenticates frames without encrypting them. If someone offers you an A on MAVLink alone, ask which criterion they think they are earning.

Who does the work

Ian Brighouse, founder and principal consultant.

Every engagement is delivered by me, not handed to a junior. I build my own tools, publish my research and write every report myself. Alongside BrighouseSec I am completing a BSc (Hons) in Cybersecurity and Networks at the University of East London.

What you will not get: inflated claims, borrowed logos or certifications I do not hold. What you will get: a fixed price, a clear deadline and work you can check. Where a measurement does not establish something, the report says so instead of scoring it.

01Own tooling. Open-source Python toolkit for passive DNS, SPF/DKIM/DMARC, TLS and HTTP header checks. Every external finding is reproducible.
02Machine learning security. RT-MLIDS: an ensemble intrusion detection framework (Random Forest + XGBoost) tested against black-box adversarial attacks, with SHAP explanations. Preprint published.
03Offensive lab work. Full Active Directory domain compromise in a lab — Kerberoasting, DCSync, Golden Ticket — mapped to MITRE ATT&CK with a CVSS-scored report.
04Clear written output. Plain-English reports a founder can act on and a buyer can read. Evidence, limitations and priorities stated, never padded.

How engagements work

Four steps. No surprises.

01

Free 20-minute call

You explain the deal, the questionnaire or the aircraft. I tell you honestly whether I can help and what it will cost.

02

Written scope and authority

A one-page scope with price, deadline and what is excluded. Nothing is tested without signed authority.

03

Delivery

Five working days on the startup side; the measurement work runs to the timetable in the scope. Drafts, findings and raw evidence together.

04

30 days of follow-up

If your buyer or your auditor comes back with questions, I help you answer them at no extra cost.

Two things people ask for that are not on this page, because they are not real yet. Checks inside your CI pipeline: the tooling runs by hand today and is not packaged for a release pipeline, so there is nothing to sell you. Remote ID broadcast testing: the Bluetooth and Wi-Fi side needs a receiver and an RF enclosure I do not own. I can measure the MAVLink path between an autopilot and a Remote ID module, and review the rest on paper, and the report will say which is which. Both will appear here when they exist.

Contact BrighouseSec

Tell me what is blocking the deal.

The questionnaire, the automation, the aircraft, or the work you want to hand off. I reply within one working day with an honest answer about whether I can help.

A free command-link check is on the table for drone makers: send me the stack and the link layout and I will tell you what I would expect to find, before anybody signs anything.

When you submit this form, I use the details to respond to your request. The form is handled by Formspree; see the Privacy Notice. Do not send passwords, credentials, private keys, payment data or sensitive personal information. An enquiry does not create a client relationship. Scope, authority, pricing and delivery terms are agreed in writing before work begins.