Akira decryptor and responder tools

File	Purpose
`akira_decrypt.py`	The decryptor (Python 3.8+, standard library only). Commands: `decrypt` (with `--keys`, `--seeds`, or the operator’s `--private-key`; `--journal` for interrupted `.arika` files), `recover` (rebuild keys from timestamps, no ransom key), `info`, `benchmark`, `selftest`
`akira_fast.py`	Optional compiled search engine for `recover` (`pip install numpy numba`). It checks itself against the Python reference at load and falls back to Python if any check fails. Takes a file from about 30 min to about 25 s
`collect_clock_and_times.ps1`	Responder collector (read-only). Records QPC/system-time pairs, a file listing with SHA-256 and timestamps, and the operator logs. Run `-Tag post` on the victim on the same boot as the attack, before any reboot

SHA-256:

akira_decrypt.py 091701682454a75dfb57324e29a6270977186f68209a7c3a93178ab76fe9fc4b. This is the 2026-10-14 version with the PDF stage-2 fallback (recover --stage2). The blind runs used 870c6db9…, the same tool without the fallback; the case folders keep that hash in their frozen logs.
akira_fast.py 07352777bff09eb30f14a8c826c503e9247948d1385b7dcc8cd02288f4606927
collect_clock_and_times.ps1 15fbf5794497b64ac954ba86bc012776971105a6a37a21137d2c4dd45eee7ff4

akira_fast.py and the collector are byte-identical to the files the blind tests ran.

Recovering a file without keys

On the victim, same boot: run collect_clock_and_times.ps1 -Tag post. From the PAIR lines, take the median of filetime − (qpc_before + qpc_after)/2, skipping the first pair. That value is --qpc-offset. Each .akira file’s last-write FILETIME is --mtime.
Known bytes at offset 0: PNG 89504e470d0a1a0a0000000d49484452, PDF 255044462d312e, JFIF JPEG ffd8ffe000104a46494600. For DOCX/XLSX/ZIP, rebuild the first local header from the central directory, which Spot mode leaves unencrypted. run6_blind_recover.py in analysis/…/04-dynamic/run6-blind/ does this per file type, and runs both passes for a whole folder.
Pass 1: python akira_decrypt.py recover f.akira --qpc-offset … --mtime … --known-hex … --out f.
Pass 2, only if pass 1 misses: add --anchor iv2 --mtime-window 2.5:0.1 --gap 8001:30000.
Large PDFs: nothing extra is needed. When the generic ChaCha8 check finds nothing in a .pdf, recover retries the same window with the PDF checks: exact object headers from the xref, then continuing a page-content stream’s deflate. This took about 2–6 minutes per file in pure Python. --stage2 pdf runs only the PDF checks; --stage2 heuristic turns them off.

Limits:

the victim must not have been rebooted, or you need a memory image;
the timestamps must survive;
at least 7 known bytes are needed at offset 0;
for files with ChaCha8 chunks (Spot files over about 262 KB, Full-mode databases, Part files over 2 MB), the ChaCha8 key search needs zeros or text in the first ChaCha chunk. PDFs have the fallback above, which needs either objects inside the ChaCha8 chunks or a page-content stream at the chunk boundary. Other compressed files come back partly decrypted.