TL;DR
- Schlarp reverse-engineered an Insta360 Link webcam, an ASUS ROG Swift PG42UQ monitor, a Shure MV7 microphone, an Elgato Cam Link 4K and an Elgato Key Light Mini, using about 13 hours of Claude working time and 98 prompts across the five.
- The headline result is the Insta360’s green recording LED, which he patched out of the firmware’s LED-pattern table while the camera carried on recording.
- The practical consequence is about cost, not capability. Per-device firmware investigation used to carry a high fixed price in specialist hours. That price has dropped, which changes how many products are worth someone’s attention.
- The self-replicating firmware worm Schlarp describes at the end of his post is his speculation. He did not build one, and no published work demonstrates one.
In plain English
Firmware is the software burned into a device that makes it behave like that device. A webcam’s recording light is not wired directly to the sensor; it is an instruction the firmware carries out when the camera enters a recording state. Change the instruction and the light stops reporting.
Whether you can change it depends on the device’s firmware integrity controls: whether it checks a cryptographic signature before accepting new firmware, whether it re-checks at boot, and whether the update path can be reached without the owner noticing. Reverse engineering is the work of finding all that out from the outside, without documentation.
Chaz Schlarp spent two weeks of evenings pointing Claude Opus 5 at five peripherals sitting within arm’s reach of his desk, and came away with a plaintext command shell inside his microphone, a webcam whose activity LED he can switch off while it records, and a key light that hands out memory writes to anyone on the Wi-Fi. He published the results on 23 August 2026.
The green light was a table entry
Schlarp’s method was the same for each device. Download the firmware and update tool from the manufacturer, drop it into his reverse-engineering environment, tell Claude Opus 5 what the goals were, and let it run. The brief asked it to reverse-engineer the update format and protocol, implement an update utility, determine the security properties of the update path including checksums and signature validation and secure boot, enumerate the protocol surfaces through static and dynamic analysis, and find hidden or debug functionality.
The Insta360 Link is the result people will repeat.
Schlarp reports that the camera runs a full ThreadX RTOS from the Ambarella SoC vendor, hosting small vision models for face tracking and gesture detection. A USB Video Class extension command kicks it into mass-storage mode, which stages a firmware update onto the device’s internal filesystem. A second vendor USB command channel offers arbitrary file read/write plus a reboot, which removes the need for the owner to replug anything. Once an image is in place, he says, there is effectively no anti-tamper — just an appended MD5 hash for integrity.
The indicator LED turned out to have a structured table of patterns in the firmware, indexed by device state. Claude wrote a tool to patch out the table entry for camera activity, repair the hash, and flash it. Schlarp’s test: the green LED that normally lights while recording no longer came on.
He is careful about the boundaries, and so should anyone repeating this. He modified firmware on his own camera, with the hardware in his hands. This is not a remote compromise of Insta360 webcams in the field. He also notes the Link’s gimbal deflects downward when it is not recording, so a second physical indicator survives the patch.
The reason it lands anyway is that the failure is so easy to follow. A recording light feels like hardware. It is an instruction, and instructions are editable when the firmware trust chain is thin.
Thirteen hours is a real number attached to a lot of human work
“An agent reverse-engineered five devices in thirteen hours” is a good headline and a poor description.
Schlarp published the breakdown himself, pulled from his Claude Code session transcripts. “Churn” is the time Claude was actually working, with long idle gaps removed. “Prompts from me” counts every message he typed, including the one-word ones telling it to keep going.
| Device | Claude churn | Prompts from Schlarp |
|---|---|---|
| Insta360 Link webcam | 3.7 hours | 33 |
| ASUS ROG Swift PG42UQ monitor | 1.2 hours | 13 |
| Shure MV7 microphone | 4.2 hours | 32 |
| Elgato Cam Link 4K | 1.5 hours | 10 |
| Elgato Key Light Mini | 2.4 hours | 10 |
| Total | 13.0 hours | 98 |
Churn is not wall-clock time and it is not an API bill. Spread across two weeks of evenings, 98 prompts is a person sitting with the work.
That brief also presumes a great deal. Someone had to know that update formats, debug functions, HID command channels, checksum routines and secure boot are the places to look. Someone had to judge whether an answer was plausible, pick the next step, notice when a generated tool might brick an expensive monitor, and confirm a result against live hardware. Schlarp did not flash the ASUS monitor at all, on the grounds that it cost a lot and he had not yet worked up the nerve.
The Cam Link is the closest thing here to unattended work: he started the run before going to sleep and woke up to a teardown and a working firmware updater. The Insta360 took 33 prompts.
So this is not push-button firmware exploitation. It is a skilled researcher using an agent to absorb the dull, repetitive, context-heavy middle of reverse engineering. That reading makes the result more useful, not less, because it tells you which part of the cost actually moved.
The microphone is the one that should worry you
The webcam LED is the memorable result. The Shure MV7 is the one with a nastier reach.
Schlarp says the firmware was hidden inside Shure’s Windows MOTIV Mix software, so Claude installed that under Wine, found the update server and pulled it down. The update protocol turned out to run over a USB HID vendor-class protocol implementing a full plaintext command shell with 48 commands: a dozen DSP controls, arbitrary memory read/write, LED control, and a four-tier privilege system whose entire authentication is a string comparison against the name of the tier you asked for. su sup just works.
The top tier can disable the touch panel so you cannot mute at the device, and drive the mute LED independently of whether the microphone is actually muted. Schlarp’s own summary: it is the webcam LED trick again, on a microphone.
Because the shell runs over HID, it is reachable from a web page in Chrome over WebHID, and Schlarp built a browser interface to demonstrate it. That removes the physical-access caveat entirely. He draws the conclusion directly: WebUSB, WebHID and WebBluetooth mean that for some devices, a moment of user indiscretion in accepting a permissions prompt could permanently backdoor an attached device.
That is a materially different threat model from “someone had your camera on a desk.” A permission dialog is a lower bar than physical possession, and it is one most people clear without reading.
Signed firmware, checked in one place
The Elgato Key Light Mini is a better lesson than the devices with no meaningful verification at all, because it is the only one of the five that had any.
Elgato signs firmware updates with Ed25519 over a SHA-512 hash of the payload and rejects images that do not validate. That is a real control, and Schlarp says the threat model justifies it: the light joins a Wi-Fi network and then offers unauthenticated access to anyone else on it.
On his account the signature protects the firmware at exactly one moment — while an update is happening. He reports no boot-time check enforced by the bootloader and no secure boot, with the updater running while the rest of the device is still operating. Schlarp asked Claude to find a way to interfere with validation, and says it found an HTTP POST that drops a payload straight into the internal UART, including a memory-poke command. He reports that one request turns the signature check into a no-op, and that he tested it with a benign patch changing the device name. None of this has been confirmed by Elgato.
This is why “we sign our firmware” is an incomplete answer to a security questionnaire. A signature does its job when the device enforces it across every path an attacker can reach. If a network-accessible service can rewrite the running updater’s memory mid-update, the cryptography worked and the system still failed. Firmware integrity is a property of the whole design: authenticated management paths, protected debug interfaces, hardened update handling, and verification at boot.
The operational lesson is smaller and more immediate. A device that accepts unauthenticated commands from anything on the same network does not belong on a network with guests, unmanaged endpoints, or anything else you have not vouched for. Same subnet is not an authentication method.
The worm is a forecast
Schlarp closes by imagining an AI-equipped worm that probes an infected host’s peripherals, relays reconnaissance to a smart command-and-control, works out how to push itself into adjacent accessories and IoT and industrial equipment, and spreads. Two things have held that back, he argues: every device model needs its own reverse engineering, and validating any of it needs the hardware in hand. The first is the labour he just handed to an agent. The second is free to malware already sitting on an infected host. He says he would not be surprised if it already exists.
That is a well-reasoned concern from someone who just did the relevant work. It is not a demonstrated capability. He did not publish a worm, and nothing in the post shows autonomous malware discovering arbitrary peripherals, building reliable exploits, surviving failures, persisting across architectures and propagating on real networks. “I wouldn’t be surprised” is professional intuition, and it should be read as intuition.
There is published evidence for the narrower claim. The 2026 FirmAgent paper pairs fuzzing with LLM agents to reconstruct vulnerability paths, run context-aware taint analysis and refine fuzzer output into proof-of-concept test cases. Evaluated on 14 real-world IoT firmware images, its authors report 182 vulnerabilities at 91% precision, including 140 previously unknown, 17 of which received CVE numbers.
That establishes growing automation in firmware vulnerability discovery. It does not establish a self-propagating firmware worm loose against consumer hardware.
The distinction is worth holding, because the unglamorous version of this risk is already the serious one. Cheaper investigation of insecure peripherals helps owners, repairers and interoperability work — Schlarp is explicit that he now has better control and understanding of his own machine. It equally helps anyone who wants a quieter place to hide on a target’s desk. Nobody needs to wait for the science-fiction version to act on the boring one.
The rules are arriving after the hardware
The EU Cyber Resilience Act is the relevant instrument for this category, and it is not yet a stick anyone can wave at these specific devices.
The Cyber Resilience Act covers products with digital elements and imposes cybersecurity obligations across planning, design, development, maintenance and vulnerability handling for the product lifecycle. It entered into force on 10 December 2024. Reporting obligations apply from 11 September 2026, and the main obligations from 11 December 2027.
A connected light or webcam sold into the EU is likely to fall inside the broad class the regulation addresses, subject to actual product classification and statutory exclusions. Its value here is not a “must sign firmware” checkbox. It pushes manufacturers toward a continuing duty to design, maintain and remediate connected products, which is precisely the gap the Key Light exposes: a signature at one stage of an update cannot make up for an unauthenticated network service that can reach into the running system.
The UK’s Product Security and Telecommunications Infrastructure regime is narrower and already in force. It has applied to relevant consumer connectable products since 29 April 2024, and requires manufacturers to ban universal and easily guessable default passwords, publish information on how to report security issues, and publish minimum security update periods.
Sensible minimums, and nowhere near a requirement for secure boot, runtime firmware verification or hardened update paths. A Wi-Fi device like the Key Light Mini may sit in scope when supplied to UK consumers. A USB-only peripheral is a more awkward fit under rules written around network-connectable products.
What I would actually do
I would not replace every webcam, microphone and light on the strength of one researcher’s post. Schlarp says he shared everything with the vendors involved. Searching for a public advisory, CVE, patch or formal statement from Shure, Insta360 or Elgato covering these specific findings, I found none as of 28 August 2026 — which means their remediation status is unknown, not that they have done nothing.
The proportionate response is duller:
- Put Wi-Fi peripherals on a network you trust, and stop treating a shared or guest network as inherently harmless.
- Apply firmware updates, and ask whether the vendor documents an update support period and a way to report security issues. Under UK rules they are supposed to publish both.
- Treat a hardware privacy indicator as one signal inside a system that itself needs protecting, rather than as proof of anything.
- Think twice about WebUSB and WebHID permission prompts. That dialog can be the whole attack.
- For organisations: put peripherals in the asset inventory and the procurement questionnaire. A webcam, monitor, capture device or microphone can carry firmware, an update channel, a network service and credentials.
- Ask vendors whether firmware integrity is enforced at boot, how debug access is controlled, whether local management interfaces require authentication, and how customers are told about security fixes.
The Insta360 LED is unsettling because it turns a familiar green light into an ordinary row in a firmware table. The devices did not become insecure this month. They were already carrying software, update mechanisms and undocumented commands. What changed is how much work it takes to find out.