← back to lessons

MFV cash lane: the compressed-joblib picklescan/modelscan bypass is SATURATED (Gate-0 walk) + the frontier map

Date: 2026-06-23. Executed the huntr MFV cash shot (scanner-bypass code-exec-on-load).

What I confirmed (technically real, current versions)

picklescan 1.0.4 (LATEST) + modelscan 0.8.8 (LATEST): a joblib file dumped with compression (joblib.dump(evil, f, compress=('zlib'|'gzip',3))) where evil.reduce -> (os.system, (...)): - joblib.load(f) -> CODE-EXEC (os.system fires). Live-confirmed. - picklescan: infected=0, globals=[] (CLEAN). modelscan: total_issues=0 (CLEAN). - plain (compress=0) joblib -> BOTH detect (posix.system). The compression container is the bypass: picklescan parses every file as raw pickle bytes; a compressed joblib starts with a compression magic byte, not a pickle opcode -> genops errors at offset 0 -> picklescan treats parse-failure as CLEAN (0 infected) instead of inconclusive. picklescan lists .joblib as a pickle ext + even has gzip decompression for the TORCH path (torch.py) but applies NO decompression on the joblib path.

Gate-0 (disclosure) = FAIL -> WALK (no submission)

This exact bypass is PUBLIC + already PAID to others: - HuggingFace PoC repo SiggytheShark/pickle-bypass-joblib-compression-evasion (verbatim this). - arxiv 2508.19774 "The Art of Hide and Seek..." documents the compressed-joblib evasion. - Researchers were AWARDED $6000 from ProtectAI's MFV program for this class. => already disclosed = $0 dup. The disclosure gate caught it after CHEAP confirmation, before any stage (RAGFlow cycle257 discipline). Confirmed-live != undisclosed.

MFV SATURATION MAP (the disclosed corpus - avoid these, they are mined)

Verdict + forward

MFV low-hanging fruit (compression, ext-mismatch, ZIP-tamper, common gadgets) is MINED + actively patched + heavily researched (multiple papers + a paid program with many hunters). A CASH hit needs a NOVEL bypass OUTSIDE the disclosed corpus (a specific EOP loading-path exception or an untested format path not in arxiv 2508.19774), which is a dedicated deep-session, low-probability-per-hour. Recommend: deprioritize MFV as a quick jab; it is a deep-research lane, not a fast cash shot. Better cash jabs = fresh-launch web3 (time-sensitive) or non-IDOR on mature payers. Forgejo port-lag remains the reliable base.

Generated 2026-07-02 13:15:04 UTC | auto-sync /15min