igneum-network/igneum
1
0
Fork 0
Code Issues Pull requests Releases Activity

attack-pass: AP-H2, a rebuild on the shared binary path reached two runs in progress; the launcher now runs from a copy in the run's scratch dir

Browse source 
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
igneum-labs 2026-10-07 15:00:26 +00:00
parent 635efc0d37
commit 405d3d5274
1 changed files with 10 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
docs/analysis/attack-pass-2026-10.md
View file

@ -256,6 +256,16 @@ check is owed to the build-server lane: the clean line should spare a lane's dec
style, or read a per-worktree exclude list), and a CI check should fail a remote-run.sh whose clean line lacks it.
OPEN until that check lands (CLAUDE.md: a rule row closes only with its check).
AP-H2 (this lane's own, 7 October 2026, 13:3x and 14:5x UK, twice). Two census runs launched from the same crate's
`target/release` binary path on the box mirror: a rebuild of the crate at a new commit replaces the binary under a
run still in progress, and every chunk the run launches after that executes the new commit's code with the old run's
label (the 07a809a7 control's later chunks ran 8bdcbdd8; the ddacfbd3 class check's later chunks ran 017e7037). Both
runs were caught by their attempt histograms (attempts 32 and 35 under a cap of 32) and their contaminated chunks
discarded. Fix in the lane's launcher: `run-census-chain.sh` copies the binary into the run's own scratch directory
before the first chunk and runs from the copy, so a rebuild cannot reach a run in progress; a run's record names the
sha256 of the copy. Class check owed: the same rule for every lane's long run (the box's build runner could refuse to
replace a binary that a running process has open, or stamp the commit into the run's log at every chunk).
## Ledger rows
AP-F1-1 (hash lane; ruling asked). At 10^5 class v4 programs one program (`attack-f1/37341`) compresses by 5.078