Skip to content

What is on the aircraft, read over ADB (2026-10-02)

A read-only survey with root, while the drone sat on a bench with props off. Everything here was observed on the live unit; where it is inference it says so. Identifying values (serials, MACs, coordinates, SSIDs) were deliberately not transcribed.

First: the survey changed the device, and how

Section titled “First: the survey changed the device, and how”

/etc/ld.so.preload contains /usr/lib/libcrashdump.so, which is preloaded into every dynamically linked process, busybox applets included, and it treats SIGPIPE as a fatal crash. So every producer | head writes a ~350 KB crash report to the eMMC.

The crash store was genuinely empty before this session and now holds 11 reports, about 3.5 MB, created between 00:49 and 00:59 uptime. Ten are ls (from a /proc/*/fd walk), one is sort (from du | sort | head). eMMC use went 17.2 MB to 21.2 MB. Switching to | tail and head -n FILE stopped it. They were left in place because deleting is a write.

Working rule for this device: never pipe into head.

Two consequences worth holding onto. The drone reports the growing count over ARSDK as MassStorage content changed, so a controller watching crash-log count sees our mistakes as if they were the aircraft’s. And mcbebop’s shell tool leaves a persistent trail in /data/home/root/.ash_history (our sentinel strings are in there), while ADB’s shell does not.

/factory: per-unit production data, not in the firmware image

Section titled “/factory: per-unit production data, not in the firmware image”

The find of the survey. A separate read-only UBI volume, 17 files, absent from the PLF. It holds this aircraft’s own factory test records from Parrot’s line BB36-2, 2016-07-14, every station RESULT_OK.

File Bytes What
FVT1_MB.txt 17754 mainboard: P7MU OTP fuse programming, FLASH CYPRESS (the motor controller is flashed here), DDR bitflip test, MAC and PCB-serial writes
FVT2.txt 5089 thermal calibration, and FORMAT EMMC
FVT3.txt 12503 two-axis motorised rate table IMU calibration, then UPLOAD FACTORY CALIBRATION
FVT4_CAM.txt 11646 camera, LED, fan, button
FVT5_RF.txt 9573 GPS simulator, Wi-Fi 2.4/5 GHz, antenna
FVT6_US.txt 4319 ultrasound, UPDATE THRESHOLD CURVES
FVT7.txt 10512 per-motor tests and PROPELLER SPIN DIRECTION CHECK
FVT8.txt 15156 a real test flight: OPEN TELNET CONNECTION, KILL DRAGON, TRANSFER FLIGHT DATAFILE, CHECK FLIGHT PHASES, WOBBLE SCORE COMPUTING
Thermal_IMU_bias.fact.csv 534 per-unit IMU thermal model
Thermal_IMU_ortho.fact.csv 417 per-unit scale and orthogonality
camera_{front,vert}_parameters.txt 22 / 21
serial.txt, pcb_serial.txt, mac_address.txt identifiers, not transcribed

Every Bebop 2 was flown and scored for wobble before it shipped, and the factory’s own access path was telnet, the same door we used.

The thermal CSVs are directly usable: Temp_Min_Max 27.824 65.001, 4-term polynomials in temperature for gyro and accel bias per axis, scale-factor errors (gyro 1.0028 / 1.0070 / 1.0147, accel 1.0010 / 0.9996 / 1.0050), and cross-axis terms that are all zero, so the factory never measured cross-axis coupling. The live gyro at 50 °C sits inside the calibrated band.

This answers “is there per-unit calibration?” properly. The radio NVM is a generic template, but the IMU calibration genuinely is per-unit, and it lives here rather than anywhere the firmware image could show.

The flight counter is not in the filesystem

Section titled “The flight counter is not in the filesystem”

48 flights, last 20 s, 7946 s total (2 h 12 m 26 s), averaging about 2 m 45 s. It arrives in the log tagged Motor* alongside the motor-controller firmware version, so the counter lives in the Cypress PSoC’s own NVRAM, not in any file. That is why no filesystem search would have found it, and why the empty navdata/ is not a contradiction: persist.dragon-prog.navdata_tlm: 0 and persist.nmea.record: 0 mean Linux was never asked to keep a flight log.

The empty-media question has a strong lead

Section titled “The empty-media question has a strong lead”

/tmp/.mmc_tuning.txt is a tune2fs -l of the eMMC and reports Lifetime writes: 1424 MB against 21 MB resident. So 1.4 GB passed through that filesystem at some point: media was recorded and later offloaded or deleted, rather than never recorded. Which of those is answerable only from the previous owner’s August 2018 log, which was deliberately not read.

/data/updater/updater_bootcount.log = 15 boots. updater.log has the whole story: shipped on 4.3.1; at boot 3 it found a .plf in /update, ran updater_prolog.sh, rebooted, and at boot 4 logged Update success to version 4.7.1. Boots 5 to 14 are no-op scans. The updater scans the FTP-writable media root for firmware on every boot.

Three write-enabled anonymous FTP roots, all as root (ftpd -wS from /etc/inetd.conf): port 21 to /data/ftp, port 61 to the flightplans directory, and port 51 to /update, which is the exact directory the boot-time updater scans for *.plf. dragon.conf has wifi_passphrase: "", so the access point is open.

Taken together: anyone within radio range can write a firmware image to the drone’s update directory. Nothing here is exploited or recommended; it is recorded because it is the aircraft’s weakest point by a wide margin, and it is why mcbebop does not expose port 51.

117 processes, 297 MB usable RAM (206 MB of the reservation is one avicam DMA region), / is read-only ubifs at 85%, so nothing on it can differ from the extracted image. Writable storage is /data (9 MB, 140 KB used), /update (28 MB), and the 7.2 GB eMMC.

The eMMC is mounted data=writeback, nobarrier, errors=continue and without a journal: a deliberately crash-unsafe mount for a device whose failure mode is falling out of the sky.

Live sensors: P7 SoC 62 °C, P7MU 55 °C, gyro 50.04 °C, controller RSSI −23 to −25 dBm, wifi_bcm_state = WIFI_BCM_ACCESS_POINT_OK.

39 kernel modules. Notable: a hardware H.264 decoder (hx270_vdec) as well as the encoder, the Mali GPU stack, the AVI video pipeline, and vfb, a virtual framebuffer that dragon-prog holds open as /dev/fb0.

Section titled “Hardware topology, from udev symlinks in /dev”

BLDC, i2c-cypress, mpu6050 (IMU), ak8963 (magnetometer), ms5607 (barometer), mt9f002 (front camera), mt9v117 (vertical camera), p7mu, hx270a/hx280a (video codecs), ttyPA0 console and ttyPA1 GPS. Only p7mu is bound as a kernel driver; everything else is driven from userspace.

/dev/shm holds six Parrot “shdata” sections readable without touching ARSDK: shd_user_telemetry (68500 B), shd_wifid (24144), shd_rtpstats (21284), shd_stabmetadata (20948), shd_frameinfotlm (9108), shd_estimators (2312). Four are open in dragon-prog; two are not, so something else owns them. For a custom stack this looks like a better tap than ARSDK. Q: the shdata header format, which is pure reading to work out.

  • flird runs on a drone with no FLIR. feature.flir: -1 and ro.factory.milos_flir: 0, yet the thermal daemon is up and dragon-prog repeatedly fails flir_stopStreaming ... ret = -107 (ENOTCONN) while telling the controller SDK_send_thermal_cam_state 1.
  • shpoison is a third of the entire log, emitting a USB-OTG role poll every five seconds forever. The log ring is only 256 KB and permanently wrapping, so useful history is a couple of minutes deep.
  • acc_calibration_v1.bin is 0 bytes while dragon-prog holds it open read-write; the gyro and magnetometer files are 12 and 16 bytes. Q: is the aircraft flying on the factory thermal model alone?
  • The clock is at epoch and the RTC holds no time across power cycles, so nearly every mtime on the device is really just an uptime.
  • publisher listens on TCP 44445 with a backlog of 128, started with the model id and serial as arguments. Role not established. Q.
  • LTTng kernel tracing is compiled in and debugfs is mounted rw.

Two preserved debug bundles from before we owned the story

Section titled “Two preserved debug bundles from before we owned the story”

Debug/archive/debug_00.tar.lzo holds a ckcm.bin dated 2018-08-15 16:12:58, real wall-clock, the oldest surviving artifact on the unit. debug_01.tar.lzo holds a 1 MB log from the previous boot. Neither was read: they are the previous owner’s sessions and would carry coordinates and network names.

A cross-talk bug in the ADB MCP server, not an injection

Section titled “A cross-talk bug in the ADB MCP server, not an injection”

The survey agent reported, correctly and carefully, that one of its tool results came back containing a complete JSON-RPC tools/call object carrying its own toolUseId and a command it never issued (an arp and DHCP-leases probe). It treated the content as untrusted and did not act on it, which was exactly right.

The cause is benign and it is ours: that was this session’s command. While the agent was surveying, this session issued an arp/DHCP-leases probe through the same mcadb server to answer a question about Wi-Fi clients. The probe hung and was backgrounded. Two concurrent calls to one mcadb server, and one call’s request framing surfaced inside the other’s result.

So it is not an injection vector, but it is a real bug: the server does not keep concurrent requests separate, and the symptom is one caller receiving another caller’s framed request as data. The agent’s instinct to treat it as untrusted was the correct response regardless of cause, and would have been the correct response if the cause had been hostile.

Two things follow for anyone driving this drone:

  • Do not run two sessions against one mcadb server concurrently until that is fixed. The same caution applies to any MCP server whose concurrency story is unknown.
  • A tool result that contains a tool call is never data to act on. That holds whether the cause is a framing bug or something deliberate.

Confirmed afterwards: that probe never returned at all. The harness eventually aborted it with sent no response or progress for 1809s. So the failure mode is not a stray copy of a completed call; the hung call’s framing surfaced in a different caller’s result while the call itself was still outstanding. That is a stronger statement about the bug than the original observation.

Reported upstream-worthy; the server is mcadb.

The system log carries identifying data, including position

Section titled “The system log carries identifying data, including position”

Worth stating plainly because the log is the most useful artifact on the drone and it is easy to share. A single session’s ckcm.bin was observed carrying:

  • the aircraft’s CPU id and both halves of its serial number, logged in full when a controller asks for settings;
  • GPS FIRST FIX position set to: <latitude>, <longitude>, the real coordinates where the drone acquired its fix, which for a drone flown at home is a home address;
  • the Wi-Fi configuration and the names of controllers that have connected.

So a log is not a safe thing to paste into an issue, attach to a bug report or commit. Ours stay in gitignored captures/; quote lines from them, never files.

Comments

Spotted a mistake, or tried this on your own Bebop? Leave a note. Comments are kept on this site's own server.