Skip to content

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.

  • 0xFF in CONFIG2 means locked, not unlocked. On the M480/M483 family, CONFIG2[15:8] (SBLOCK) and CONFIG2[7:0] (ALOCK) use 0x5A as their explicit unlock token1. Any other byte value - including an erased or default 0xFF - activates the lock.
  • M483SIDAE has no hardware AES accelerator or XOM, but SBLOCK still blocks ICE reads. The M483xIDAE sub-family omits the cryptographic accelerator and Execute-Only Memory present on GCAE/E8AE parts2. When SBLOCK != 0x5A, Note 6 of the TRM ICE Lock Effect Table (Table 6.5-7) marks every ICE-initiated FMC operation except MassErase invalid and sets the ISP Fail Flag (ISPFF)14.
  • A repeated DPIDR word (0x2ba01477) at memory addresses is a failure signature, not a memory read. Reading CPUID (0xe000ed00), CONFIG0..CONFIG3 (0x00300000), APROM (0x00000000), and peripheral registers (0x50080100, 0x50080110) over raw Nu-Link HLA commands returned 0x2ba01477 at 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 8,651 trigger=none pulses with no synchronization to a target reset or boot event (and multi-second OpenOCD startup latency after a 35-37 us reset response), treating an update-mode application ciphertext backup (65 pages, 66,560 bytes) 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.

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.

Flash Initialization ControllerReloads User Configurationat POR or active resetCONFIG0 (LOCK, ICELOCK, MBS, CBS)CONFIG2 (SBLOCK[15:8], ALOCK[7:0])Comparator gateUnlocked only when:LOCK == 1 ANDALOCK == 0x5A ANDSBLOCK == 0x5Alatched stateCPU master path(TRM Table 6.5-6)APROM: R/W allowedBoot Loader (0x00800000): R allowedOTP/SBK (0x00300500): masked 0xA5internal busSWD / ICE master path(TRM Table 6.5-7 Note 6)When SBLOCK != 0x5A:all ops except MassErase invalidFMC sets ISPFF = 1AHB-AP / HLA debugObserved SWD readbackAll-0xFF or repeated DPIDR(0x2ba01477 at every address)No address sensitivityblocked read

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 observationWhat it genuinely provesWhat it does not proveMinimum gate to advance
SWD returns DPIDR = 0x2ba01477Physical SWD clock/data continuity to the Cortex-M4 Debug Port (DP)That AHB-AP memory reads work, or which M483 sub-variant is fittedRead 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 wordsThat CONFIG0 or CONFIG2 actually contains 0xFFFFFFFF in flashQualify the read path against an unlocked M480/M483 control board
10 us crowbar pulse drops Faultier ADC min from 0.388 to 0.0235The 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 thresholdMeasure nanosecond waveforms on an oscilloscope (>= 100 MHz)
40 us VDD pulse drives nRESET low 35-37 us after VDD fallExternal reset assertion occurs above 35-40 us of VDD droop on this harnessWhether shorter pulses trigger internal POR/BOD/LVR, or when CONFIG latchesSynchronise 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 dataThat the file matches the currently wired PCB, or restores LDROM/OTPBind backup manifest to a hardware board_id and verify round-trip restore

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 KEYPRT hardware key-protection registers).
  • No XOM (Execute-Only Memory) region support.
  • Standard FMC readout-protection and boot locks stored in the User Configuration block starting at 0x00300000 (CONFIG0 through CONFIG3, with extended configuration words up to CONFIG10).

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 bitsField nameErased / default valueUnlocked valueDocumented behaviour in M480 TRM Rev. 3.07
CONFIG0[1]LOCK11 (0 = locked)Flash content is unlocked only if LOCK == 1, ALOCK == 0x5A, and SBLOCK == 0x5A1
CONFIG0[5]MBS11 (APROM/LDROM)Master Boot Selection; 0 or SBLOCK != 0x5A routes boot through the mask-ROM Boot Loader1
CONFIG0[7:6]CBS11Application-definedChip Boot Selection between APROM and LDROM when MBS == 1
CONFIG0[11]ICELOCK11 (0 = ICE disabled)Controls whether the Cortex-M4 debug interface (ICE) is enabled
CONFIG2[15:8]SBLOCK0xFF0x5ASecure Boot Lock; any value other than 0x5A locks flash and restricts ICE to Chip Erase1
CONFIG2[7:0]ALOCK0xFF0x5AAdvance Security Lock; any value other than 0x5A enables security lock even when LOCK == 11

Two structural consequences follow from this layout:

  1. Both SBLOCK and ALOCK live in the same 32-bit word (CONFIG2 at 0x00300008). For the FMC to treat flash as unlocked over SWD, both comparators must evaluate to 0x5A (CONFIG2[15:0] == 0x5A5A) while CONFIG0[1] (LOCK) is 1 and CONFIG0[11] (ICELOCK) is 1. Flipping a single bit in CONFIG0 cannot unlock a chip whose SBLOCK or ALOCK byte is 0xFF.
  2. An all-0xFF SWD read of 0x00300000 is not a reading of CONFIG0..CONFIG3. Early in the investigation (2026-11-01), an all-0xFF SWD read of the CONFIG block was interpreted literally as “the chip’s CONFIG words are 0xFFFFFFFF, so SBLOCK = 0xFF and ALOCK = 0xFF”. While SBLOCK != 0x5A is consistent with the official NuMicro ICP tool reporting Target flash is in security lock mode, a locked chip returning 0xFFFFFFFF (or 0x2ba01477) at every address is returning a bus-level lock mask, not the stored flash bytes. You cannot know the exact programmed CONFIG bits 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 / operationAddress windowCPU master (Table 6.5-6, SBLOCK active)ICE / SWD master (Table 6.5-7 + Note 6, SBLOCK != 0x5A)
APROM (512 KB application flash)0x00000000-0x0007FFFFR W (CPU can read and write APROM)Blocked (Note 6 overrides column; sets ISPFF = 1)
LDROM (4 KB loader flash)0x00100000-0x00100FFFBlocked or restricted by LOCK/SBLOCKBlocked (Note 6: only MassErase valid)
User Configuration (CONFIG0..3)0x00300000-0x0030000FReadable by CPU / boot pathTable column says R W, but Note 6 + live SWD block reads
OTP / Secure Boot Key (SBK)0x00300500-0x003005FFDirect reads return masked 0xA5 (0xA5A5A5A5)Blocked (no read/compare oracle before unlock)
Mask-ROM Boot Loader (32 KB)0x00800000-0x00807FFFR - (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-initiated FMC operation sets the ISP Fail Flag (ISPFF, bit 6 / 0x40 in FMC_ISPSTS and FMC_ISPCTL)4. The TRM’s ISPFF condition 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.c flash driver in OpenOCD5 does not attempt ISP reads on a locked chip; its unlock path writes the CHEAT register and triggers a whole-chip mass erase.
  • As for the Secure Boot Key (SBK) region in OTP (0x00300500-0x003005FF) and the adjacent configuration page around 0x00300800, static analysis of the vendor boot path showed that direct reads return a constant 0xA5 (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:

  1. Address sensitivity. Querying two different addresses that contain different values must return two different words.
  2. Failure-mode separation. The classifier must distinguish a transport error or adapter buffer echo from a valid bus read.
  3. 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:

  1. Standard OpenOCD mdw failed at target examination. Before running mdw 0x00300000 1, OpenOCD’s Cortex-M target layer reads the architectural CPUID register at 0xe000ed00. The trace log showed that reading 0xe000ed00 returned 0x2ba01477 - the exact DPIDR value. OpenOCD extracted bits [15:4] as PARTNO = 0x147 (unrecognised Cortex-M part), left the target unexamined, and rejected mdw with Target not examined yet.
  2. Bypassing CPUID examination via raw_mdw returned 0x2ba01477 everywhere. Adding a custom HLA command (raw_mdw <address>) that called the Nu-Link memory-read API directly bypassed OpenOCD’s CPUID gate. Querying the locked chip then produced:
    • CONFIG0..CONFIG3 (0x00300000..0x0030000C): 0x2ba01477 at every word.
    • First 16 words of APROM (0x00000000..0x0000003C): 0x2ba01477 at every word.
    • Peripheral register addresses (0x50080100, 0x50080110): 0x2ba01477 at 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 0x2ba01477 at 0xe000ed00, 0x00300000, 0x00000000, and 0x50080100 has zero address sensitivity.
  • Because two similarly locked OP1 boards both returned 0x2ba01477 (and earlier 0xFFFFFFFF), the bench had no unlocked M483 positive control showing that raw_mdw over the Nu-Link HLA transport would return real flash words even if a glitch briefly cleared the FMC lock latch.
  • Worse, a repeated DPIDR value on an HLA adapter is indistinguishable from the adapter’s USB firmware returning its last valid SWD buffer word when an AHB-AP read stalls or faults (just as disconnected IDCODE reads echoed the 0x0b9f ADC voltage word twice as 0x0b9f0b9f).

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 identifierTrigger modeRequested attemptsCompleted rowsRecorded outcome classesFinal state
Initial LDO_CAP grid (2026-11-14)nRESET rising edge1,5061,5061,506 all-ff / locked-responseExhausted, disarmed=true
Baseline VDD grid (2026-11-15)trigger=none (OUT_NONE)1,5061,5061,506 locked-responseExhausted, disarmed=true
Wide VDD grid (2026-11-15)trigger=none (OUT_NONE)10,0007,1457,142 locked-response, 3 readback_errorStopped 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=none and OUT_NONE, the target MCU was already booted and idling in USB application/update mode. Varying Faultier’s delay parameter 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 a POR or reset-release CONFIG-reload window.
  • Five-order-of-magnitude readback latency mismatch. As the logic-analyzer pulse ladder in the companion harness guide showed, a VDD crowbar pulse of 40 us or longer causes nRESET to assert 35-37 us after the VDD falling edge. Even if a pulse triggered a reset and transiently corrupted a volatile configuration latch, the runner checked the result by spawning a fresh openocd OS 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 sourceApplication image sizeApp SHA-256 prefixConfiguration image sizeConfiguration SHA-256 prefix
Board A pre-update backup66,560 B (65 pages)33b06cf01,024 B71809669
Board B in-boot backup (2026-11-14)66,560 B (65 pages)5194d9381,024 B71809669
Board C bring-up backup (2026-11-15)66,560 B (65 pages)33b06cf01,024 Bc1917858

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:

  1. 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 KB LDROM bootloader (0x00100000), the FMC CONFIG0..CONFIG3 words (0x00300000), or OTP (0x00300500). If a voltage glitch during flash initialisation corrupts LDROM or CONFIG, an application-partition backup cannot unbrick the board.
  2. Equal hashes do not prove board identity or fresh capture. Board A and Board C shared the exact same 33b06cf0 application hash, while Board B differed (5194d938), and Board A and Board B shared the 71809669 configuration hash. A hash proves file integrity on disk, not which physical PCB is currently under the needles.
  3. Readback is not write-restoration. Proving that 65 pages 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 ns per sample, any pulse between 4 ns and 128 ns falls inside a single ADC conversion window.
  • With the factory 2.2 uF decoupling capacitor still soldered to LDO_CAP (pin 50), nanosecond crowbar pulses are absorbed by the capacitor’s charge reservoir; with the crowbar on VDD (pin 51), the onboard LDO regulator and bulk decoupling sit between the crowbar drain and the 1.26 V internal core domain (LDO_CAP) where the FMC sense amplifiers run6.
  • Without an oscilloscope trace on the core rail, there is no evidence that a sub-microsecond pulse on VDD reaches the internal FMC comparators 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.

1. Non-glitch CPU/ISP path exhausted?Check TRM CPU vs ICE table (Table 6.5-6 vs 6.5-7).Probe read-only mask-ROM / UART / USB ISP first.2. Positive readback oracle qualified?Does the exact readback command return distinct,address-sensitive words on an unlocked control target?ISP absent / lockedSTOP: do not launch a timing grid.More attempts cannot fix an unqualified oracleor an unsynchronised trigger.Untested ISP route exists3. Target event and readback latency matched?Is the trigger tied to a measured POR or nRESET edge,and can readback run before the transient state clears?Yes (address-sensitive)No (all-0xFF or repeated DPIDR)4. Pulse waveform verified at the core rail?Does an oscilloscope or time-resolved capture showdroop at the actual pulse width under test?Yes (edge-coupled)No (trigger=none / slow spawn)5. Board-bound recovery scope accepted?Is the backup bound to this board_id, and is theunbacked risk (LDROM / CONFIG) explicitly accepted?YesNo (sub-sample ADC only)Run bounded, logged calibration ladder(stop on first anomaly; verify address sensitivitybefore calling any changed word plaintext)YesNo
  1. Exhaust non-glitch CPU/ISP routes first. Because Table 6.5-6 allows the CPU to read APROM even when SBLOCK != 0x5A1, check whether the mask-ROM Boot Loader (0x00800000) or LDROM exposes a read-only command over USB (0416:a316 / 0416:3f00) or UART1 (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; UART1 remains a separate read-only check.)
  2. Qualify an address-sensitive positive control. Never start a sweep whose preflight accepts locked-response (0x2ba01477) or all-ff as proof that memory reads work. Verify the readback transport (such as persistent vendor-interface SWD over Faultier interface 1 or OpenOCD raw_mdw) against an unlocked M480/M483 board first.
  3. Couple the trigger and readback to a measured target event. Never run a delay grid with trigger=none. Tie the glitch timer to a physical POR ramp or nRESET edge, and use an in-process SWD readback rather than spawning a multi-second OS process per trial.
  4. Measure pulse shape at the timescale under test. Treat Faultier’s 2.604 MS/s ADC and a 1-2 MHz logic analyzer as tools for microsecond (>= 1 us) delivery and reset-threshold checks; require an oscilloscope before sweeping nanosecond (4-128 ns) widths.
  5. Bind recovery artifacts to the physical board and scope. Record the target’s board_id in the backup manifest, verify round-trip restoration, and treat LDROM/CONFIG corruption as an explicit unbacked risk.
Claim or observationStatusHow it was checked
M483SIDAE (M483xIDAE) has no hardware AES/crypto engine and no XOMDocumentedNuvoton M480 Series Datasheet Rev. 3.09 selection table2
CONFIG2 SBLOCK[15:8] and ALOCK[7:0] require 0x5A to unlock; 0xFF is lockedDocumentedNuvoton 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 ISPFFDocumentedNuvoton 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 resetDocumentedNuvoton 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 valuesMeasured (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_mdwMeasured (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 readsMeasured (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 highMeasured (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)
PracticeEvidence
Require an address-sensitive positive control before any glitch sweepraw_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 configurationsWithout 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 sessionnRESET 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 plaintextA 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 regionsTwo 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).
  1. 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, and ISPFF conditions 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

  2. Nuvoton Technology Corp., “NuMicro Family M480 Series Datasheet,” Rev. 3.09, product selection tables (M483xIDAE vs M483GCAE/E8AE) and electrical characteristics. https://www.nuvoton.com/export/resource-files/en-us—DS_M480_Series_EN_Rev3.09.pdf ↩ ↩2 ↩3 ↩4 ↩5

  3. xPack Dev Tools, “xPack OpenOCD,” commit 7bb05570f40dae2c742e6e581db4129468fed689. https://github.com/xpack-dev-tools/openocd ↩

  4. OpenNuvoton contributors, “fmc_reg.h,” M480BSP CMSIS device header (FMC_ISPSTS_ISPFF_Msk and FMC_ISPCTL definitions). https://github.com/OpenNuvoton/M480BSP/blob/master/Library/Device/Nuvoton/M480/Include/fmc_reg.h ↩ ↩2 ↩3

  5. OpenNuvoton contributors, “numicro.c,” OpenOCD-Nuvoton NOR flash driver. https://github.com/OpenNuvoton/OpenOCD-Nuvoton/blob/master/src/flash/nor/numicro.c ↩

  6. 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 ↩

  7. Nuvoton Technology Corp., “NuMicro M480 Series Secure Boot Application Note (AN0039),” mask-ROM bootloader SBLOCK and HSUSBD/UART1 command-mode behaviour. https://www.nuvoton.com/export/resource-files/en-us—TRM_M480_Series_EN_Rev3.07.pdf ↩