aboutsummaryrefslogtreecommitdiffstats
path: root/hw/ta_handlefuzz.py (follow)
Commit message (Collapse)AuthorAge
* ninety-fourth pass: handle and state-machine fuzz, negative but weakuser2026-09-30
Payload fuzzing cannot reach lifetime bugs, and the defects found statically earlier in this project are the other shape, where a handle from one command is replayed into another including after destruction. A generic driver for that shape harvests the dwords each command writes as candidate handles and replays them, as-is and mutated, into every command, flagging a handle accepted where a random value in the same field is rejected. Four applications show no such acceptance, but this is weak evidence and is recorded as such. Handles are injected at a single fixed offset so a real handle field elsewhere is never exercised, only a handful of candidates were harvested per application with one application yielding a single candidate, and if a command ignores that field then the control is also accepted so the discriminator never fires. The known lifetime defects were found by reading the slot life-cycle rather than by blind replay, which is consistent with this generic version being too coarse to rediscover them. A meaningful version would need each application's actual handle field offsets. The objective is not met and this class is explicitly not counted as covered.