aboutsummaryrefslogtreecommitdiffstats
path: root/hw/ta_cmdfuzz.py (follow)
Commit message (Collapse)AuthorAge
* seventy-sixth pass: extended command surface clean, root cause still openuser2026-09-30
| | | | | | | | | | | | | | | | | | | | Command ids twelve to thirty-one across six applications with four payload profiles each are entirely clean, with no odd statuses, stalls or state drift. Combined with the previous pass the silicon coverage is twelve applications over the lower command range and six over the extended range, with exactly two faults, both reaching the same system-driver opcode. The root cause of that opcode's hang is still not established and is deliberately not guessed at. The worker's first callee bounds its index, indexes a fixed-stride table and performs a chain of crypto-shaped calls, with no blocking loop visible at that level, and that callee in turn dispatches into the small-opcode handlers, so the hang is at least three levels deep. The chain has already been mis-read twice, including an attribution falsified on hardware, so no third hypothesis is recorded without a test behind it. The objective is not met. Hardware access produced a real reproducible host-triggered defect in the platform root of trust, but a denial of service is not a memory-corruption primitive and the one attempt to connect the two was falsified by the first experiment that tested it.
* seventy-second pass: hardware fuzz finds a host INVOKE that wedges the PSPuser2026-09-30
With the card authorised, the twelve-application command surface was fuzzed on silicon across command ids and payload profiles with a fixed re-probe after each command to catch state drift. One application's command six wedges the platform security processor: a ten second stall, then every later ring submit returns an incrementing error, no further application loads, and re-opening the device fails with a power-management timeout. The eleven load failures later in that run are this wedge rather than separate faults. It was isolated and reproduced in a clean run where the two preceding commands return instantly as in-run controls, and a zeroed payload is sufficient with no crafted input. Statically the handler validates its selector and then issues a service call with an opcode in the upper range, so it lands in the large dispatcher and the hang is in the system driver rather than the application. The objective is deliberately still not met: this is a new hardware-confirmed host-reachable defect but it is a denial of service, not a demonstrated memory-corruption primitive, since a stall followed by a dead ring is equally consistent with a poll loop on an unpowered block. The classifying experiment is named in the finding. The card was recovered after each wedge and verified healthy.