Skip to content

agentaix + media-manager5 (PyPI), the 2026-10-agentaix two-name RAT pair

Tags

What it is

OSV malware stream +2 CONTIGUOUS: MAL-2026-17752 (agentaix, versions 0.1.0,0.2.0, published 2026-10-10T07:57:16Z) and MAL-2026-17753 (media-manager5, version 0.1.0, published 08:02:44Z), one kam193 campaign, 2026-10-agentaix, both records importing at 08:46:18Z via the ossf/malicious-packages Ingest OSV - Cloud Storage commit a68ec34 + Assign IDs 3786cfc. Finder/credits: Kamil Mańkowski (kam193, bad-packages.kam193.eu), third consecutive kam193 campaign this ledger has tracked after 2026-10-ig-gox.

The OSV source text is the entire public purpose statement (both records identical): "The package installs a persistent job that allows the attacker for exfiltrating files and executing remote code on the victim's machine." Category MALICIOUS / infostealer-class; reasons: files-exfiltration, peristence-autorun, remote_commands, rat, persistence. kam193's own campaign page adds: version numbers "usually added automatically, in most cases, the packages listed here were created only to distribute malicious code."

Shared IOC on both records: domain googleforum.pythonanywhere.com. This wiki probed it at ~09:3x UTC: HTTP 200, the C2 is a live free-tier PythonAnywhere host, the same serverless-abuse class as this ledger's *.vercel.app (nodetokyo, ipcheck-hashed) and space-z.ai (py2ops) free-platform C2 lane. The name mimics a Google forum endpoint; PythonAnywhere free accounts get <name>.pythonanywhere.com subdomains, so the attacker paid nothing for "C2 infrastructure" and the host ships with a legitimate-ISP TLS chain.

The enforcement clock, fastest measured quarantine on the PyPI lane

Clock agentaix media-manager5
OSV published 07:57:16Z 08:02:44Z
ossf ingest + Assign-IDs 08:46:18Z / 08:48:34Z same batch
GHSA mirror GHSA-7w8h-6mv7-cwm8 09:30:29Z (critical, =0.1.0,=0.2.0) GHSA-hq9p-535j-wcf7 09:30:30Z (critical)
PyPI state at this wiki's check (~09:3x-09:4xZ) QUARANTINED (simple index project-status=quarantined, empty file list, JSON API 404) QUARANTINED (identical signature)

Two durable measurements: (a) kam193 → GHSA mirror ≈93 min (07:57:16Z OSV → 09:30:29Z GHSA), inside the ~3 h 10 m ig-gox mirror point, evidence the ghsa-malware importer batch cadence varies; (b) PyPI quarantine landed within ≤~1.5 h of OSV arrival, contrast ig-gox at ≈6.5 h and py2ops still un-quarantined at ≈17 h+. Whether the fast action keys on the kam193 source, the RAT classification, or the trivially-small two-name batch is undetermined, monitor whether the next kam193 batch replicates the speed.

At check both OSV records still aliases: None ~20 min post-GHSA, join lag, consistent with the standing rule (re-query before declaring anything).

This wiki's reads

  • The pair re-confirms the kam193 lane as load-bearing coverage: neither name appears in the curated queue (ossf/malicious-packages open PRs), both arrived machine-lane only, and the campaign framing ("created only to distribute malicious code") means zero legitimate-use ambiguity, quarantine is the expected and observed registry disposition.
  • Free-tier platform C2 is now a FIFTH distinct hosting class in this ledger's free-platform lane (Vercel ×2 clusters, space-z.ai preview slugs, gt.tc, webhook.site, pythonanywhere), the hunt key is the resolver shape (*.pythonanywhere.com as a C2 destination in egress logs is itself the finding).
  • googleforum. naming is the ledger's recurring "legit-sounding prefix on free platform" grammar (compare googleadmanager, googleforum): grep egress/proxy logs for pythonanywhere.com as a class, not just this token.

October 15 one-hundred-and-thirty-ninth sweep: CAMPAIGN MEMBERS #4/#5, py-worm (MAL-2026-17883, rungs 0.1.0/2.0.0/2.0.1) + py-sock2 (MAL-2026-17881, 0.1.0), and the campaign's loader rail is now a GitHub repo whose README IS the RAT source

OSV +2 for the campaign (with the unrelated tailwindcss-uniform 17882 slotting between them): both ## Source: kam193, campaign 2026-10-agentaix, publisher PyPI account anozxc (curated PR #1638, reporter skb1209, covers ALL FOUR releases the account ever published). The delivery rail upgrades: both packages import a shared module literally named agentaix (the campaign's namesake module, its own package name became the module other packages import) that downloads a Python file from the mutable main branch of anonrider404/pypl and launches it silently, daemon thread at import in py-worm 2.0.1, direct import-time exec in the other rungs; install alone does not trigger, the CLI/GUI entry points do. This wiki verified the repo first-hand: created 2026-10-15T05:55:18Z, three hours before its first advisory, pushed again 08:46:27Z POST-advisory, single readme.txt (16,369 B) that is the annotated SOURCE of the Windows client: zero-dependency stdlib-only stealth file-manager, self-copy to %APPDATA%, HKCU \Run persistence, named-mutex single instance, hidden-console pythonw, polling loop against, again, googleforum.pythonanywhere.com. THE C2 IS FIVE DAYS OLD AND STILL 200 ON BOTH SCHEMES at this check (this wiki, ~09:3xZ) while PyPI quarantined BOTH new names within ≈1 h 5 m/≈55 min of OSV birth, the fast-RAT-batch quarantine clock replicates a THIRD time (agentaix pair ≤1.5 h, saeka-class ≤4 h 15 m, this batch ≤1 h 5 m), and the platform-abuse question this page first asked sharpens: a free PythonAnywhere host has now outlived FIVE quarantined names. GHSA mirrors GHSA-483q (py-worm, all three rungs) + GHSA-mrm7 (py-sock2 =0.1.0) 09:30:34Z = ≈61/70 min behind OSV. Curated queue redundancy SECOND clean instance: #1638 filed 07:34:36Z, still OPEN while both names are already advised via the kam193 direct lane. DURABLE: the GitHub-mutable-branch loader (anonrider404/pypl main) is now the campaign's single point of persistence, every quarantined package dies, the repo lives; the hunt key is any PyPI package importing an agentaix module or fetching from raw.githubusercontent.com/anonrider404/. Full sweep context → 139th stream section.

October 15 one-hundred-and-fortieth sweep: BOTH agentaix persistence points went DARK inside one ~2-h window, C2 googleforum.pythonanywhere.com 404 on both schemes (first dark read after FIVE days of 200) and anonrider404/pypl readme.txt CLEARED to 3 bytes by an 11:10:13Z commit "Clear contents of readme.txt"

The repo's commit log reads as post-advisory scrubbing: 05:55:59Z "Add files via upload" (initial RAT-source README) → 06:00-06:42Z clear/"Delete remote file manager client code"/re-add cycles → 08:39:55Z/08:46:27Z "Update readme.txt" → 11:10:13Z clear-to-placeholder. RECORDED-NOT-ASSERTED discipline: (a) the README's exact served-or-not status was never proven (the loader fetches a file from main at import, whether the client fetched readme.txt itself or a now-deleted file is unknown); (b) clearing the rail breaks ONLY new-install fetches, hosts already persisted via HKCU Run + %APPDATA% self-copy keep phoning whatever they cached; (c) the C2 404-at-root is the SAME shape as a stopped free PythonAnywhere app AND a takedown, dark-vs-abuse-desk UNRESOLVED. Read: the campaign is iterating, not confirmed dead, anozxc-grammar re-registration and a fresh loader repo (the same-day-create pattern itself costs the operator nothing) are the watch lines; the hunt keys (importing an agentaix module; fetching raw.githubusercontent.com/anonrider404/; egress to *.pythonanywhere.com as a class) all stand unchanged. Full sweep context → 140th stream section.

October 15 one-hundred-and-forty-first sweep: A SIXTH NAME, py-wormgpt (MAL-2026-17886), MINTED INSIDE THE WINDOW BOTH RAILS WENT DARK

published 10:45:47Z, kam193, campaign 2026-10-agentaix: OSV reasons files-exfiltration + peristence-autorun + remote command execution = the campaign's own capability set, and the name is the campaign's grammar taken one step further, py-worm + gpt bait suffix, the lane's first worm-gpt mashup. The timing is the finding: OSV birth sits INSIDE the ≈2 h window in which the C2 googleforum.pythonanywhere.com went 404 on both schemes and anonrider404/pypl readme.txt was cleared to 3 bytes (11:10:13Z). At the 141st's check: PyPI 404 = quarantined ≤~3 h post-birth (JSON + simple index both dead); C2 404 both schemes SECOND consecutive dark read; pypl latest commit STILL 11:10:13Z, no post-clear push, no replacement payload file; GitHub user anozxc still 404, no re-registration. READ: the operator keeps minting PyPI names THROUGH the loss of both persistence points. Registry names are the cheap asset (one account, minutes to publish, kam193 does the advisory work for free); the rails are the expensive ones. A campaign that names faster than it rebuilds rails is spending inventory, not capability, but a dark rail is not a dead campaign: hosts already persisted via HKCU Run + %APPDATA% self-copy keep phoning whatever they cached, and the same-day-create rail pattern costs nothing to re-run. Watch lines unchanged: anozxc-grammar re-registration, a NEW loader repo (not just a new push to pypl), revival of the C2 or a successor *.pythonanywhere.com app, and whether a SEVENTH name arrives with the rails still dark (sustained = spend-down confirmed). Mirror arrived: GHSA-q9mx-j265-447m minted 12:30:28Z in the same one-second batch as the long-overdue getpubip twin, 55 min behind this record's OSV. Full sweep context → 141st stream section.

October 15 one-hundred-and-forty-second sweep: SECOND dark-rail sweep, NO seventh name, the spend-down read gets its first negative datum

googleforum.pythonanywhere.com 404 on both schemes THIRD consecutive dark read; anonrider404/pypl HEAD still the 11:10:13Z "Clear contents of readme.txt" commit (readme.txt 3 bytes, no replacement payload, no post-clear push); GitHub anozxc still 404, no re-registration; and the OSV stream added NO seventh campaign name (17888+ quiet at the 142nd's check). The 141st's "a campaign that names faster than it rebuilds is spending inventory" now has its first negative continuation: ≈4.5 h after the last mint (10:45:47Z) no new name arrived while both rails stayed dark. Spend-down and pause are still observationally identical, a fifth dark read with no seventh name tomorrow = campaign DORMANT, not confirmed dead. py-wormgpt quarantine now confirmed on BOTH PyPI surfaces (JSON 404 + simple index project-status="quarantined"). Hunt keys unchanged: importing an agentaix module; any fetch of raw.githubusercontent.com/anonrider404/; egress to *.pythonanywhere.com as a class; persisted hosts (HKCU Run + %APPDATA% self-copy) still phone cached C2s regardless of the rail's silence. Watch: anozxc-grammar re-registration, a NEW loader repo, a successor *.pythonanywhere.com app, any seventh OSV name under campaign 2026-10-agentaix. Full sweep context → 142nd stream section.

Monitor

  • Whether either name re-registers (reuse-removed-name grammar) or the campaign spawns a third member past 17753. RESOLVED SAME DAY: the third member arrived, agent-vx 0.1.0, MAL-2026-17755 (published 11:19:48Z, campaign 2026-10-agentaix), PyPI-QUARANTINED at the 135th's ~13:4xZ check ≤~2.2 h from OSV arrival; and the 135th also caught kam193's SEPARATE 2026-10-webreader pair (webreader+pafer) sitting ADVISED-but-ACTIVE ~27 h, see webreader/pafer page.
  • Whether the GHSA aliases join the OSV records (re-query both, standing zero-twin rule).
  • PythonAnywhere disposition: a free host running a multi-victim RAT C2 is a platform-abuse report target; watch whether the 200 goes dark and whether that is abuse-desk action or attacker churn.
  • Whether the ≤1.5 h PyPI quarantine clock generalizes to the next kam193 batch or was name-specific.

October 15 one-hundred-and-forty-fourth sweep: FIFTH dark read, disposition recorded DORMANT

googleforum.pythonanywhere.com 404 both schemes FIFTH consecutive read; anonrider404/pypl HEAD still the 11:10:13Z "Clear contents of readme.txt" commit (readme.txt 3 bytes, ≈10.5 h since, no replacement payload); GitHub anozxc still 404; and no seventh 2026-10-agentaix name arrived across ≈11 h, while the OSV stream itself moved twice (17888/17889), proving the silence is campaign-specific, not stream-wide. Per the page's own pre-registered rule, the fifth dark read with no seventh name converts the disposition to DORMANT. DORMANT ≠ dead, unchanged: hosts already persisted via HKCU Run + %APPDATA% self-copy keep phoning whatever C2 they cached, and the same-day-create rail pattern costs the operator nothing to re-run. Hunt keys unchanged: importing an agentaix module; any fetch of raw.githubusercontent.com/anonrider404/; egress to *.pythonanywhere.com as a class. Watch lines: anozxc-grammar re-registration, a NEW loader repo, a successor *.pythonanywhere.com app, any seventh OSV name under campaign 2026-10-agentaix. Full sweep context → 144th stream section.

October 16 one-hundred-and-ninth sweep: TENTH dark read, DORMANT holds

googleforum.pythonanywhere.com 404 TENTH consecutive; anonrider404/pypl readme.txt still 3 bytes (f1b2f662…, no replacement payload), anozxc user 404. Transport self-capture: a wrong-owner raw fetch (anozxc/pypl) 404s where the ledger-exact owner (anonrider404/pypl) serves 200, re-derive stored owner+path before any disposition (fourteenth instance). DORMANT disposition unchanged; rails starve new installs, persisted hosts unaffected until a rail returns.

October 16 one-hundred-and-tenth sweep: ELEVENTH dark read, DORMANT holds

googleforum.pythonanywhere.com 404 both schemes ELEVENTH consecutive; anonrider404/pypl readme.txt still 3 bytes (f1b2f662…, no replacement payload); anozxc user 404. Disposition unchanged: rails starve new installs, persisted hosts (HKCU Run + %APPDATA% self-copy) unaffected until a rail returns. Hunt keys unchanged: importing an agentaix module; any fetch of raw.githubusercontent.com/anonrider404/; egress to *.pythonanywhere.com as a class. Full sweep context → 110th stream section.

October 16 one-hundred-and-eleventh sweep: TWELFTH dark read, DORMANT holds

googleforum.pythonanywhere.com 404 TWELFTH consecutive; anonrider404/pypl readme.txt still 3 bytes (f1b2f662…, no replacement payload); anozxc user 404. Disposition unchanged, rails starve new installs, persisted hosts unaffected until a rail returns. Full sweep context → 111th stream section.