Workstations and toolingreference
Nuvoton M483 flash locks and the glitch-oracle trap
Non-destructive readout of a locked microcontroller by voltage-fault injection is an instrumentation problem before it is a timing problem. This reference documents the Flash Memory Controller (FMC) protection model on the Nuvoton M480/M483 family (M483SIDAE)12 and the methodological post-mortem of an 8,651-attempt trigger=none (10,157 total attempts including 1,506 reset-rising trials) fault-injection investigation on my own Endgame Gear OP1 8k v2 mouse - where the lock held, no key or plaintext APROM was extracted over SWD, and a post-run review of the journals showed that the sweep had been firing against a repeatable failure signature rather than a qualified memory-read oracle.
For the physical wiring, single-rail-owner power sequencing that prevents Faultier RP2040 back-powering (Windows error code 10), and the patched OpenOCD driver used during these tests, see the companion guide Read-only SWD and voltage-glitch harness on a locked Nuvoton M483. For the broader workflow combining static binary decompilation with black-box protocol captures prior to hardware probing, see Agent-assisted reverse engineering.
0xFFinCONFIG2means locked, not unlocked. On theM480/M483family,CONFIG2[15:8](SBLOCK) andCONFIG2[7:0](ALOCK) use0x5Aas their explicit unlock token1. Any other byte value - including an erased or default0xFF- activates the lock.M483SIDAEhas no hardware AES accelerator orXOM, butSBLOCKstill blocks ICE reads. TheM483xIDAEsub-family omits the cryptographic accelerator and Execute-Only Memory present onGCAE/E8AEparts2. WhenSBLOCK != 0x5A, Note 6 of the TRM ICE Lock Effect Table (Table 6.5-7) marks every ICE-initiatedFMCoperation exceptMassEraseinvalid and sets the ISP Fail Flag (ISPFF)14.- A repeated
DPIDRword (0x2ba01477) at memory addresses is a failure signature, not a memory read. ReadingCPUID(0xe000ed00),CONFIG0..CONFIG3(0x00300000),APROM(0x00000000), and peripheral registers (0x50080100,0x50080110) over raw Nu-Link HLA commands returned0x2ba01477at every word. Treating that repeatable response as a working readback oracle left the sweep unable to distinguish chip-level lock enforcement from adapter-level stale buffer echoes. - Three other evidence promotions compounded the blind oracle. Running
8,651trigger=nonepulses with no synchronization to a target reset or boot event (and multi-second OpenOCD startup latency after a35-37 usreset response), treating an update-mode application ciphertext backup (65pages,66,560bytes) as proof of full-chip recoverability, and treating a microsecond ADC rail dip (0.388 -> 0.0235) as proof that nanosecond (4-128 ns) pulses reached the core latch all masked missing prerequisites.
FMC protection topology and access paths
Section titled “FMC protection topology and access paths”The M480/M483 Flash Memory Controller evaluates protection differently depending on whether the bus master is the internal Cortex-M4 CPU (Table 6.5-6) or an external SWD/ICE debug probe (Table 6.5-7)1.
Early diagnostic table: what each observation actually proves
Section titled “Early diagnostic table: what each observation actually proves”When probing a locked microcontroller, most early milestones prove less than they appear to. Use this table to separate what a measurement confirms from the stronger claim it is often promoted to support.
| Bench observation | What it genuinely proves | What it does not prove | Minimum gate to advance |
|---|---|---|---|
SWD returns DPIDR = 0x2ba01477 | Physical SWD clock/data continuity to the Cortex-M4 Debug Port (DP) | That AHB-AP memory reads work, or which M483 sub-variant is fitted | Read an address-dependent register (CPUID, SYS_PDID) and verify distinct values |
Raw SWD read returns 0xFFFFFFFF or 0x2ba01477 at 0x00300000 (CONFIG0) | The read path did not return valid plaintext configuration words | That CONFIG0 or CONFIG2 actually contains 0xFFFFFFFF in flash | Qualify the read path against an unlocked M480/M483 control board |
10 us crowbar pulse drops Faultier ADC min from 0.388 to 0.0235 | The MOSFET pulls the probed rail down at 10 us (~26 ADC samples at 2.604 MS/s) | That 4-128 ns pulses cross the core LDO or FMC sense threshold | Measure nanosecond waveforms on an oscilloscope (>= 100 MHz) |
40 us VDD pulse drives nRESET low 35-37 us after VDD fall | External reset assertion occurs above 35-40 us of VDD droop on this harness | Whether shorter pulses trigger internal POR/BOD/LVR, or when CONFIG latches | Synchronise readback to nRESET release with sub-millisecond latency |
USB update-mode backup saves 65 pages (66,560 B app + 1,024 B configuration) | The bootloader exports encrypted application pages and configuration data | That the file matches the currently wired PCB, or restores LDROM/OTP | Bind backup manifest to a hardware board_id and verify round-trip restore |
The Nuvoton M480/M483 flash-lock model
Section titled “The Nuvoton M480/M483 flash-lock model”What M483SIDAE includes and omits
Section titled “What M483SIDAE includes and omits”The Nuvoton M480 family spans parts with full cryptographic secure boot (M487, M483GCAE, M483E8AE) and general-purpose high-speed USB parts (M483SIDAE in the M483xIDAE selection group)2. According to the Datasheet Rev. 3.09 selection table, M483SIDAE has:
- No hardware cryptography peripheral (no AES, DES, SHA, or ECC engine, and no
KEYPRThardware key-protection registers). - No
XOM(Execute-Only Memory) region support. - Standard
FMCreadout-protection and boot locks stored in the User Configuration block starting at0x00300000(CONFIG0throughCONFIG3, with extended configuration words up toCONFIG10).
Eliminating hardware AES and XOM narrows the problem: any protection on M483SIDAE comes from the standard FMC configuration gates and mask-ROM / LDROM boot flow rather than a dedicated cryptographic coprocessor.
CONFIG0 and CONFIG2: why 0xFF means locked
Section titled “CONFIG0 and CONFIG2: why 0xFF means locked”A common trap when reading Nuvoton documentation quickly is assuming that erased flash (0xFF) leaves every protection field open. While single-bit flags in CONFIG0 use 1 for their inactive/default state, the byte-wide lock tokens in CONFIG2 invert that intuition by requiring an explicit 0x5A magic byte to unlock1:
| Register bits | Field name | Erased / default value | Unlocked value | Documented behaviour in M480 TRM Rev. 3.07 |
|---|---|---|---|---|
CONFIG0[1] | LOCK | 1 | 1 (0 = locked) | Flash content is unlocked only if LOCK == 1, ALOCK == 0x5A, and SBLOCK == 0x5A1 |
CONFIG0[5] | MBS | 1 | 1 (APROM/LDROM) | Master Boot Selection; 0 or SBLOCK != 0x5A routes boot through the mask-ROM Boot Loader1 |
CONFIG0[7:6] | CBS | 11 | Application-defined | Chip Boot Selection between APROM and LDROM when MBS == 1 |
CONFIG0[11] | ICELOCK | 1 | 1 (0 = ICE disabled) | Controls whether the Cortex-M4 debug interface (ICE) is enabled |
CONFIG2[15:8] | SBLOCK | 0xFF | 0x5A | Secure Boot Lock; any value other than 0x5A locks flash and restricts ICE to Chip Erase1 |
CONFIG2[7:0] | ALOCK | 0xFF | 0x5A | Advance Security Lock; any value other than 0x5A enables security lock even when LOCK == 11 |
Two structural consequences follow from this layout:
- Both
SBLOCKandALOCKlive in the same 32-bit word (CONFIG2at0x00300008). For theFMCto treat flash as unlocked over SWD, both comparators must evaluate to0x5A(CONFIG2[15:0] == 0x5A5A) whileCONFIG0[1](LOCK) is1andCONFIG0[11](ICELOCK) is1. Flipping a single bit inCONFIG0cannot unlock a chip whoseSBLOCKorALOCKbyte is0xFF. - An
all-0xFFSWD read of0x00300000is not a reading ofCONFIG0..CONFIG3. Early in the investigation (2026-11-01), anall-0xFFSWD read of theCONFIGblock was interpreted literally as “the chip’sCONFIGwords are0xFFFFFFFF, soSBLOCK = 0xFFandALOCK = 0xFF”. WhileSBLOCK != 0x5Ais consistent with the official NuMicro ICP tool reportingTarget flash is in security lock mode, a locked chip returning0xFFFFFFFF(or0x2ba01477) at every address is returning a bus-level lock mask, not the stored flash bytes. You cannot know the exact programmedCONFIGbits of a locked board from a blocked SWD read.
CPU versus ICE access asymmetry and the ISPFF gate
Section titled “CPU versus ICE access asymmetry and the ISPFF gate”The M480 TRM splits flash access permissions into two separate tables: Table 6.5-6 (operations initiated by the CPU via FMC In-System Programming registers) and Table 6.5-7 (operations initiated by an external ICE/ICP/Writer tool)1.
| Memory region / operation | Address window | CPU master (Table 6.5-6, SBLOCK active) | ICE / SWD master (Table 6.5-7 + Note 6, SBLOCK != 0x5A) |
|---|---|---|---|
APROM (512 KB application flash) | 0x00000000-0x0007FFFF | R W (CPU can read and write APROM) | Blocked (Note 6 overrides column; sets ISPFF = 1) |
LDROM (4 KB loader flash) | 0x00100000-0x00100FFF | Blocked or restricted by LOCK/SBLOCK | Blocked (Note 6: only MassErase valid) |
User Configuration (CONFIG0..3) | 0x00300000-0x0030000F | Readable by CPU / boot path | Table column says R W, but Note 6 + live SWD block reads |
OTP / Secure Boot Key (SBK) | 0x00300500-0x003005FF | Direct reads return masked 0xA5 (0xA5A5A5A5) | Blocked (no read/compare oracle before unlock) |
| Mask-ROM Boot Loader (32 KB) | 0x00800000-0x00807FFF | R - (readable in all protection columns) | Blocked by ICE lock gate |
Why ICE-initiated FMC register writes fail (Note 6 and ISPFF)
Section titled “Why ICE-initiated FMC register writes fail (Note 6 and ISPFF)”At first glance, Table 6.5-6 suggests an easy workaround: since the CPU is allowed to read APROM under SBLOCK, why not use SWD to write an ISP read command (FMC_ISPCMD = 0x00, target address in FMC_ISPADR, trigger ISPGO in FMC_ISPTRG, poll FMC_ISPSTS for ISPBUSY == 0, and read FMC_ISPDAT at 0x4000C008), or halt the CPU and inject a small SRAM stub?
Note 6 of Table 6.5-7 closes all three routes: “All operations (except for MassErase) are invalid for ICE when SBLOCK is not 0x5A.”1 When SBLOCK != 0x5A:
- Any
ICE-initiatedFMCoperation sets the ISP Fail Flag (ISPFF, bit 6 /0x40inFMC_ISPSTSandFMC_ISPCTL)4. The TRM’sISPFFcondition list explicitly includes “Invalid operations with ICE connection if SBLOCK is not 0x5A” (TRM Rev. 3.07, pages 711 and 721)1. - OpenNuvoton’s own
numicro.cflash driver in OpenOCD5 does not attempt ISP reads on a locked chip; its unlock path writes theCHEATregister and triggers a whole-chip mass erase. - As for the Secure Boot Key (
SBK) region inOTP(0x00300500-0x003005FF) and the adjacent configuration page around0x00300800, static analysis of the vendor boot path showed that direct reads return a constant0xA5(0xA5A5A5A5) mask unless preceded by a matching unlock write. There is no partial-match or timing oracle on that region from outside the chip.
The oracle problem: how a failure signature masquerades as a memory read
Section titled “The oracle problem: how a failure signature masquerades as a memory read”In fault-injection testing, an oracle is the measurement that tells the runner whether a single glitch pulse changed the target’s security state. To be valid for a flash-readout sweep, a readback oracle must satisfy three properties:
- Address sensitivity. Querying two different addresses that contain different values must return two different words.
- Failure-mode separation. The classifier must distinguish a transport error or adapter buffer echo from a valid bus read.
- Positive-control verification. Before attacking the locked target, the exact readback command must be proven to return expected, address-dependent words on an unlocked control target (or an unlocked address range on the same bus).
What actually happened on the locked M483SIDAE
Section titled “What actually happened on the locked M483SIDAE”Once the physical SWCLK contact fault on pin 34 was repaired on 2026-11-14 and the safe-connect OpenOCD driver connected at 200 kHz (2980 mV sense voltage, supply=0), the probe returned the exact Cortex-M4 DPIDR (0x2ba01477) across three consecutive runs. The next step was reading memory:
- Standard OpenOCD
mdwfailed at target examination. Before runningmdw 0x00300000 1, OpenOCD’s Cortex-M target layer reads the architecturalCPUIDregister at0xe000ed00. The trace log showed that reading0xe000ed00returned0x2ba01477- the exactDPIDRvalue. OpenOCD extracted bits[15:4]asPARTNO = 0x147(unrecognised Cortex-M part), left the target unexamined, and rejectedmdwwithTarget not examined yet. - Bypassing
CPUIDexamination viaraw_mdwreturned0x2ba01477everywhere. Adding a custom HLA command (raw_mdw <address>) that called the Nu-Link memory-read API directly bypassed OpenOCD’sCPUIDgate. Querying the locked chip then produced:CONFIG0..CONFIG3(0x00300000..0x0030000C):0x2ba01477at every word.- First 16 words of
APROM(0x00000000..0x0000003C):0x2ba01477at every word. - Peripheral register addresses (
0x50080100,0x50080110):0x2ba01477at every word.
Why locked-response was not a qualified oracle
Section titled “Why locked-response was not a qualified oracle”The runner code classified any 16-word APROM read consisting entirely of 0xFFFFFFFF as all-ff and any read consisting of repeated 0x2ba01477 as locked-response. For a sample to be promoted to plaintext, the classifier required a plausible Cortex-M SRAM initial stack pointer (0x2000xxxx) in word 0 and a Thumb reset vector in APROM/LDROM in word 1, confirmed by a second identical read.
That classifier properly rejected 0x2ba01477 as plaintext, but the preflight gate accepted locked-response as proof that the readback oracle worked. That was a category error:
- Returning
0x2ba01477at0xe000ed00,0x00300000,0x00000000, and0x50080100has zero address sensitivity. - Because two similarly locked OP1 boards both returned
0x2ba01477(and earlier0xFFFFFFFF), the bench had no unlockedM483positive control showing thatraw_mdwover the Nu-Link HLA transport would return real flash words even if a glitch briefly cleared theFMClock latch. - Worse, a repeated
DPIDRvalue on an HLA adapter is indistinguishable from the adapter’s USB firmware returning its last valid SWD buffer word when anAHB-APread stalls or faults (just as disconnectedIDCODEreads echoed the0x0b9fADC voltage word twice as0x0b9f0b9f).
Three more ways evidence promotion derailed the sweep
Section titled “Three more ways evidence promotion derailed the sweep”A post-run review of the session history and run journals on 2026-11-15 identified three additional places where a scoped observation was promoted into a stronger assumption that the data did not support.
1. Timing sweeps without a qualified target event (trigger=none and readback latency)
Section titled “1. Timing sweeps without a qualified target event (trigger=none and readback latency)”Where in the boot cycle does the M483 sample CONFIG0/CONFIG2? The M480 TRM (printed page 652, Flash Initialization Controller) states: “When chip is powered on or active from reset, Flash Initialization Controller checks the flash stability first, and then reloads User Configuration to flash control registers” (with reset-source exceptions for BL/BS documented on pages 711-712)1.
After an initial 1,506-attempt reset-rising sweep (0-250 us delay, 4-128 ns width on LDO_CAP with the decoupling capacitor installed) returned no change, a working note hypothesised that SBLOCK is latched only at cold Power-On Reset (POR) and never on nRESET. Later summaries repeated that POR-only hypothesis as an established fact, even though the TRM explicitly lists “powered on or active from reset” and the actual latch window had never been measured.
Acting on that unverified premise, two larger sweeps were launched on 2026-11-15 with trigger=none and power-cycling disabled (OUT_NONE):
| Run identifier | Trigger mode | Requested attempts | Completed rows | Recorded outcome classes | Final state |
|---|---|---|---|---|---|
Initial LDO_CAP grid (2026-11-14) | nRESET rising edge | 1,506 | 1,506 | 1,506 all-ff / locked-response | Exhausted, disarmed=true |
Baseline VDD grid (2026-11-15) | trigger=none (OUT_NONE) | 1,506 | 1,506 | 1,506 locked-response | Exhausted, disarmed=true |
Wide VDD grid (2026-11-15) | trigger=none (OUT_NONE) | 10,000 | 7,145 | 7,142 locked-response, 3 readback_error | Stopped on 3rd error, disarmed=true |
Look at what the two trigger=none runs (8,651 attempts combined) actually did in hardware:
- No synchronisation to any target event. With
trigger=noneandOUT_NONE, the target MCU was already booted and idling in USB application/update mode. Varying Faultier’sdelayparameter only changed how many microseconds the RP2040 waited after receiving the host USB command before pulsing the MOSFET. Every delay value was firing into the same steady-state idle loop, never intersecting aPORor reset-releaseCONFIG-reload window. - Five-order-of-magnitude readback latency mismatch. As the logic-analyzer pulse ladder in the companion harness guide showed, a
VDDcrowbar pulse of40 usor longer causesnRESETto assert35-37 usafter theVDDfalling edge. Even if a pulse triggered a reset and transiently corrupted a volatile configuration latch, the runner checked the result by spawning a freshopenocdOS process after the pulse - a step that takes seconds on Windows, missing a microsecond boot window by five orders of magnitude.
2. Intact update-mode backups promoted to full-chip recovery proof
Section titled “2. Intact update-mode backups promoted to full-chip recovery proof”Before running hardware fault trials on a live device, a sound safety rule is to capture a verified backup so any corrupted flash page can be restored. Using the mouse’s USB update-mode HID protocol (3367:1977), a read-only backup tool dumped all 65 application pages (66,560 bytes total, starting with the expected 96 bytes of header padding across page ranges 0x00-0x33, 0x34-0x74 application ciphertext, and 0x75-0x7F) alongside a 1,024-byte configuration block.
When hashed across three physical OP1 8k v2 boards, the captures showed:
| Capture source | Application image size | App SHA-256 prefix | Configuration image size | Configuration SHA-256 prefix |
|---|---|---|---|---|
| Board A pre-update backup | 66,560 B (65 pages) | 33b06cf0 | 1,024 B | 71809669 |
| Board B in-boot backup (2026-11-14) | 66,560 B (65 pages) | 5194d938 | 1,024 B | 71809669 |
| Board C bring-up backup (2026-11-15) | 66,560 B (65 pages) | 33b06cf0 | 1,024 B | c1917858 |
The runner checked that the recovery files existed on disk and matched their recorded SHA-256 hashes before arming the crowbar. Promoting that hash check into “full recovery is proven” hid three gaps:
- Application ciphertext is not a full-chip dump. The update-mode reader exports the encrypted application partition and its configuration block. It does not back up the
4 KBLDROMbootloader (0x00100000), theFMCCONFIG0..CONFIG3words (0x00300000), orOTP(0x00300500). If a voltage glitch during flash initialisation corruptsLDROMorCONFIG, an application-partition backup cannot unbrick the board. - Equal hashes do not prove board identity or fresh capture. Board A and Board C shared the exact same
33b06cf0application hash, while Board B differed (5194d938), and Board A and Board B shared the71809669configuration hash. A hash proves file integrity on disk, not which physical PCB is currently under the needles. - Readback is not write-restoration. Proving that
65pages can be read in update mode does not prove that a standalone script can write them back and exit update mode without the vendor’s official updater binary.
3. Undersampled ADC dips promoted to sweep qualification
Section titled “3. Undersampled ADC dips promoted to sweep qualification”Faultier’s onboard ADC_CROWBAR (250 MHz / 96 = 2.604 MS/s, ~384 ns per sample) recorded a normalised baseline of ~0.388 on LDO_CAP (pin 50, 1.26 V) and ~0.919-0.933 on VDD (pin 51, 3.3 V). When a 10 us test pulse dropped the LDO_CAP minimum to 0.0235 and the VDD minimum to 0.0627, that confirmed DC continuity and microsecond MOSFET conduction.
Promoted into “the crowbar path is qualified for a nanosecond sweep”, it masked the sampling physics:
- At
384 nsper sample, any pulse between4 nsand128 nsfalls inside a single ADC conversion window. - With the factory
2.2 uFdecoupling capacitor still soldered toLDO_CAP(pin 50), nanosecond crowbar pulses are absorbed by the capacitor’s charge reservoir; with the crowbar onVDD(pin 51), the onboardLDOregulator and bulk decoupling sit between the crowbar drain and the1.26 Vinternal core domain (LDO_CAP) where theFMCsense amplifiers run6. - Without an oscilloscope trace on the core rail, there is no evidence that a sub-microsecond pulse on
VDDreaches the internalFMCcomparators before triggering a whole-chip brown-out reset.
Decision guide: qualifying a read-only fault-injection experiment
Section titled “Decision guide: qualifying a read-only fault-injection experiment”Before launching any automated glitch sweep on a locked MCU, walk through the qualification gates in order. Every gate is a hard stop if it fails.
- Exhaust non-glitch CPU/ISP routes first. Because
Table 6.5-6allows the CPU to readAPROMeven whenSBLOCK != 0x5A1, check whether the mask-ROM Boot Loader (0x00800000) orLDROMexposes a read-only command over USB (0416:a316/0416:3f00) orUART1(PF.0/PF.1, pins 33/34)7 before pulsing power rails. (On the OP1 8k v2, USB HID ISP was absent on power-up;UART1remains a separate read-only check.) - Qualify an address-sensitive positive control. Never start a sweep whose preflight accepts
locked-response(0x2ba01477) orall-ffas proof that memory reads work. Verify the readback transport (such as persistent vendor-interface SWD over Faultier interface 1 or OpenOCDraw_mdw) against an unlockedM480/M483board first. - Couple the trigger and readback to a measured target event. Never run a delay grid with
trigger=none. Tie the glitch timer to a physicalPORramp ornRESETedge, and use an in-process SWD readback rather than spawning a multi-second OS process per trial. - Measure pulse shape at the timescale under test. Treat Faultier’s
2.604 MS/sADC and a1-2 MHzlogic analyzer as tools for microsecond (>= 1 us) delivery and reset-threshold checks; require an oscilloscope before sweeping nanosecond (4-128 ns) widths. - Bind recovery artifacts to the physical board and scope. Record the target’s
board_idin the backup manifest, verify round-trip restoration, and treatLDROM/CONFIGcorruption as an explicit unbacked risk.
Evidence table
Section titled “Evidence table”| Claim or observation | Status | How it was checked |
|---|---|---|
M483SIDAE (M483xIDAE) has no hardware AES/crypto engine and no XOM | Documented | Nuvoton M480 Series Datasheet Rev. 3.09 selection table2 |
CONFIG2 SBLOCK[15:8] and ALOCK[7:0] require 0x5A to unlock; 0xFF is locked | Documented | Nuvoton M480 TRM Rev. 3.07 section 6.5 (CONFIG0/CONFIG2 descriptions, p. 659)1 |
When SBLOCK != 0x5A, all ICE operations except MassErase are invalid and set ISPFF | Documented | Nuvoton M480 TRM Rev. 3.07 Table 6.5-7 Note 6 and ISPFF conditions (pp. 711, 721)1; fmc_reg.h4 |
Flash Initialization Controller reloads CONFIG at power-on or active from reset | Documented | Nuvoton M480 TRM Rev. 3.07 printed page 652 (with BL/BS reset rules on pp. 711-712)1 |
OTP Secure Boot Key (SBK, 0x00300500-0x003005FF) reads return masked 0xA5 values | Measured (static) | Static decompilation and M480 boot-path review (2026-11-14) |
Locked M483SIDAE returns DPIDR 0x2ba01477 on CPUID (0xe000ed00), CONFIG0 (0x00300000), and APROM (0x00000000) via raw_mdw | Measured (live) | Safe-connect OpenOCD logs and manifests on OP1 8k v2 boards (2026-11-14) |
1,506 reset-rising + 8,651 trigger=none (1,506 baseline + 7,145 wide VDD) attempts produced 0 plaintext reads | Measured (live) | Recomputed complete journal.jsonl attempt records (2026-11-14 to 2026-11-15) |
VDD crowbar pulses of 40 us to 1 ms assert nRESET 35-37 us after VDD fall; 1-35 us pulses leave nRESET high | Measured (live) | fx2lafw logic analyzer captures at 1 MHz and 2 MHz paired with Faultier ADC counts (2026-11-15) |
Update-mode 65-page (66,560 B) application hashes differ across boards (33b06cf0 vs 5194d938) while 1,024 B configuration hashes can match (71809669) | Measured (live) | SHA-256 verification across three OP1 8k v2 boards (2026-11-14 to 2026-11-15) |
What to do about it
Section titled “What to do about it”| Practice | Evidence |
|---|---|
| Require an address-sensitive positive control before any glitch sweep | raw_mdw returned 0x2ba01477 at 0xe000ed00, 0x00300000, 0x00000000, and 0x50080100; accepting locked-response in preflight hid a zero-sensitivity oracle across 8,651 trigger=none rows. |
Reject trigger=none in automated timing-grid configurations | Without a hardware POR or nRESET trigger edge, varying delay only shifts when the MOSFET fires relative to the host USB command, testing 0 points of the boot-time CONFIG latch window. |
| Couple readback to the trigger over a persistent SWD session | nRESET asserts 35-37 us after a >= 40 us VDD droop, whereas spawning a fresh openocd process per trial takes seconds and misses any transient unlock state. |
| Treat changed words on a locked target as an anomaly first, never plaintext | A single non-0xFF, non-DPIDR read can come from SWD framing corruption; require a repeatable re-read and an address-dependent match across multiple offsets before claiming an unlock. |
Bind backup manifests to a hardware board_id and state unbacked regions | Two boards shared a 1,024 B configuration SHA-256 (71809669) with different 66,560 B app images (33b06cf0 vs 5194d938), and update-mode backups omit LDROM (0x00100000) and CONFIG (0x00300000). |
Related docs
Section titled “Related docs”- Read-only SWD and voltage-glitch harness on a locked Nuvoton M483 - the companion step-by-step hardware guide covering continuity isolation, single-rail-owner sequencing against RP2040 ESD diode back-powering (
Windows error code 10), the safe-connect OpenOCD patch, and the1 usto1 msVDDpulse ladder. - Agent-assisted reverse engineering - how static binary decompilation and protocol capture narrow the hardware questions before touching a PCB.
- 8 kHz mouse polling on Windows: merged HID reports - USB high-speed HID report analysis on the OP1 8k v2 and its peers.
- PCB-as-code verification - automated geometry and netlist gates for custom hardware.
- GRIDDLE LX Hall-effect keyboard - bringing up custom analog-sensing peripheral hardware and firmware.
References
Section titled “References”-
Nuvoton Technology Corp., “NuMicro Family M480 Series Technical Reference Manual,” Rev. 3.07, Dec. 2023, Flash Memory Controller (
FMC) section 6.5 (pp. 652, 659, Tables 6.5-6 and 6.5-7, Note 6, andISPFFconditions on pp. 711, 721). https://www.nuvoton.com/export/resource-files/en-us—TRM_M480_Series_EN_Rev3.07.pdf ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14 ↩15 ↩16 ↩17 ↩18 -
Nuvoton Technology Corp., “NuMicro Family M480 Series Datasheet,” Rev. 3.09, product selection tables (
M483xIDAEvsM483GCAE/E8AE) and electrical characteristics. https://www.nuvoton.com/export/resource-files/en-us—DS_M480_Series_EN_Rev3.09.pdf ↩ ↩2 ↩3 ↩4 ↩5 -
xPack Dev Tools, “xPack OpenOCD,” commit
7bb05570f40dae2c742e6e581db4129468fed689. https://github.com/xpack-dev-tools/openocd ↩ -
OpenNuvoton contributors, “fmc_reg.h,” M480BSP CMSIS device header (
FMC_ISPSTS_ISPFF_MskandFMC_ISPCTLdefinitions). https://github.com/OpenNuvoton/M480BSP/blob/master/Library/Device/Nuvoton/M480/Include/fmc_reg.h ↩ ↩2 ↩3 -
OpenNuvoton contributors, “numicro.c,” OpenOCD-Nuvoton NOR flash driver. https://github.com/OpenNuvoton/OpenOCD-Nuvoton/blob/master/src/flash/nor/numicro.c ↩
-
0xrafasec, “Voltage glitching the STM32F1 read-out protection: a step-by-step crowbar attack,” 0xrafasec Blog. https://0xrafasec.com/en/posts/voltage-glitching-the-stm32f1-read-out-protection-a-step-by-step-crowbar-attack ↩
-
Nuvoton Technology Corp., “NuMicro M480 Series Secure Boot Application Note (AN0039),” mask-ROM bootloader
SBLOCKandHSUSBD/UART1command-mode behaviour. https://www.nuvoton.com/export/resource-files/en-us—TRM_M480_Series_EN_Rev3.07.pdf ↩