Skip to content

pyyi (PyPI), the py2ops operator's NEXT NAME

Tags

What it is

Curated queue PR ossf/malicious-packages #1627 (saic12, filed 2026-10-12 10:08:35Z) files PyPI pyyi 2.4.1. This wiki's same-day verification: single release 2.4.1, uploaded 2026-10-12T03:45:51Z (wheel sha256 003828175a747a9a3de8029e6a028ad2c3739abf1dad7b468b25636e391e91f4, 9,099 B, wheel-only, no sdist), latest = 2.4.1, PyPI simple meta project-status: active, ZERO OSV (v1/query empty) + ZERO GHSA (pip affects=pyyi empty) ≈11.7 h live at this check. The curated lane caught it in ≈6.4 h; the machine lanes (kam193, amazon-inspector, this ledger's two most reliable feeders) have zero record.

The header comment is the confession, quoted verbatim: "python2, operator-linked interpreter shell (connection layer only)… nothing of value ships here. bare python2 behaves exactly like the local python interpreter - because it IS one (a real REPL); that is the whole cover. every line is read with the console echo switched off and put back on the record afterwards."

The same-operator proof (why this is a rebrand, not a copycat)

pyyi is the continuation of py2ops, the identity link is infrastructure, not style:

  1. Byte-identical baked C2 rails. DEFAULT_SERVER = https://preview-chat-960504a9-b4e0-4b69-bc80-6ac3d4fa4ae0.space-z.ai (env PY2_SERVER overrides) and FALLBACK_SERVER = https://b196q82k8891-d.space-z.ai, the SAME two slugs baked into py2ops, the same panel the ledger has watched enroll devices since Oct 9. A copycat would mint its own panel; this artifact ships the incumbent's.
  2. Version continuity across the rename. py2ops iterated 2.2 → 2.2.1 → 2.3 → 2.3.1 (Oct 9-10) then froze ≈69 h at this check. pyyi's single release is 2.4.1, the next minor step in py2ops' own sequence, as if the version line never broke. The rename preserved the release counter.
  3. Feature set = py2ops 2.3.1 + more. pyyi ships the full late-py2ops arsenal from birth: remember_word()/recall_word() (unlock word XOR-scrambled under a per-machine key in ~/.python2/state.json), _rekey() silent re-registration on 401/403, _failover() server shopping on stale rails, echo-off _quiet_read() (termios / msvcrt.getwch), the ≤5-tries bare-token unlock gate, the lone-0 quiet reply. Same config dir ~/.python2, same apiKey registration shape, same UA family python2/<ver>.
  4. NEW: the payload fetch is in the wheel. Late py2ops said "every working command is served by the linked ops server" but shipped no fetch-exec path this readable. pyyi.run_payload() does: POST /api/cli/payload {apiKey} → gzip.decompress(base64.b64decode(payload)) → exec(compile(src,"<adv>","exec")) in memory, nothing on disk, every failure path masked as python: can't open file '…': [Errno 2] No such file or directory (rc 2), the "boring cover for every failure" in the source's own words. It self-heals around locked/pinned/stale servers with retries, re-register, and re-unlock using the remembered word before ever showing the cover.
  5. python2 -d self-destruct removes ~/.python2 + copies under /usr/local/bin, ~/.local/bin, and four Windows WindowsApps/AppData paths, cleanup ergonomics, not installer hygiene.

Timeline that makes the rebrand behavioral, not coincidental: py2ops advisory PR #1613 was PUBLIC from Oct 9 16:02Z; pyyi 2.4.1 published Oct 12 03:45:51Z, before the 05:23Z merge and 05:30Z mint, i.e. the operator moved names while watching the queue, the same advisory-as-build-trigger grammar the ledger clocked on @baanx, pullgetsage, and orderly-plugin, this time triggered by a PR, not a record. py2ops' merged record is 2.2-only (false-clean certificate on live 2.3.1), and by the time it landed, the operator's live artifact was under a different name on the same panel. The advisory burned the name, not the account, not the panel, not the code, the util-foriepft lesson reproduced one ecosystem-cycle later, with the version counter as the fingerprint.

This wiki's checks (Oct 12 ~15:2x-15:3x UTC)

  • Wheel pulled, unzipped, read statically, no execution. Contents: pyyi/__init__.py (87 B), pyyi/__main__.py (118 B), pyyi/cli.py (23,923 B), dist-info. entry_points.txt: python2 = pyyi.cli:main, the console script collides by design with the interpreter name. METADATA: author pyyi, keywords python2, no homepage, no project URLs (the py2ops metadata C2 homepage leak is closed, the bake moved into cli.py).
  • POST /api/cli/init {"hostname":"probe","os":"linux"} (ledger-exact baked spellings, harmless probe): preview-chat-960504a9-… → 200 {"ok":true,"apiKey":"ptm_hrbo…"} and b196q82k8891-d → 200 {"ok":true,"apiKey":"ptm_6m1t…"}, the enrolment endpoint SURVIVED behind the login-UI flip the 156th flagged: both rails answer 307→/login at root AND still mint device keys at /api/cli/init. The panel serves UI and enrolment simultaneously; the /login front is an OPSEC coat, not a shutdown. Keys keep the ptm_ prefix.
  • Name-keyed OSV v1/query EMPTY; GHSA ecosystem=pip&affects=pyyi EMPTY ≈11.7 h. #1627 open unmerged ≈5.4 h at check, until merged, every advisory-keyed consumer is blind, and #1627's own text is accurate ("silently registers the host with a remote server… Base64-encoded, gzip-compressed code… execute in memory").
  • py2ops same-check: still 2.3.1, no 2.3.2 (release gap ≈69 h), record scope still ['2.2'], 17807 aliases still None.

Durable reads

  • The panel is the cluster key. Two package names, one baked slug pair, space-z.ai preview-chat-<uuid> + <slug>-d is now a two-artifact infrastructure signature. Any future package baking these hosts, or that slug shape, joins this lineage regardless of name, author, or description.
  • Detection preceded naming. Every pyyi hunt key is ALREADY on the ledger from py2ops: ~/.python2/config.json / state.json (key k post-first-unlock marker), UA python2/<ver>, space-z.ai registration traffic, console_scripts: python2, POSTs to /api/cli/init|unlock|payload. The rename buys the operator nothing against an artifact/traffic hunt, only against name-keyed advisory consumers. That asymmetry (advisories are name-scoped, hunts are infra-scoped) is the whole durable lesson.
  • Rename latency as an operator clock: PR-public → rebrand-publish ≈3.5 d (Oct 9 16:02Z → Oct 12 03:45Z). If py2ops-scope records ever widen to 2.3.x, expect either another name or the same name re-iterating, both are watch-answers, neither is comfort.
  • The exec-in-memory + No such file or directory failure-mask pair is a forensic-shape choice: disk-only EDR and file-integrity monitors see the config dir only; the second stage never lands. Response guidance must key process behavior (python process dialing space-z.ai, termios echo-off toggles, msvcrt raw reads in a python2-named script) not file hashes.

October 12 one-hundred-and-thirteenth sweep: the enrolment "cadence" was OURS, per-request minting confirmed inside one window

Two harmless ledger-exact POST /api/cli/init calls against b196q82k8891-d.space-z.ai minutes apart this sweep returned different keys: ptm_zp8s7xmowvcpv7f87jjguiwz, then ptm_o1t4630w2hayq3fhbmdroe2j. The endpoint mints a fresh ptm_ key per request, it is an open enrolment faucet, not a clock. Correction on the ledger's own read: the "seventh key, cadence ≈1.5/day" series across the 109th-112th sweeps was counting this wiki's own probe schedule, not operator key-rotation. Durable: a value sampled once per sweep cannot produce an operator cadence; cadence claims need within-window repeat samples first. Same-sweep adjacent capture: www.space-z.ai reads 410 and a FABRICATED control subdomain (zzznosuchrail9x7.space-z.ai) reads 410 identically = the 410 at the apex is wildcard-front catch-all, not a disposition of the real host; the load-bearing half of the 156th's read stands (device rails mint while the front looks dark: b196 root 307 + init 200 re-confirmed). preview-chat-960504a9-… root TIMED OUT (000) this check (110th 404, 111th 307), the rail's root-level reads are time-of-check only; the /api/cli/init route on b196 remains the only stable enrolment oracle.

Monitor

  • pyyi iterations (2.4.2+ / same-version re-upload under the maintainer change watch), quarantine arrival, OSV/GHSA mint clock vs the curated #1627 merge clock (which lane writes the record this time, machine, curated, or GHSA-first like 17810 did today?).
  • Any NEW package name baking either space-z.ai slug, or the ptm_ apiKey prefix on any endpoint.
  • preview-chat-<uuid> slug churn: does the baked slug rotate to match panel re-deploys (FC re-deploy is minutes-slow per the 156th's Alibaba FC read).
  • /api/cli/payload shape, the one route the ledger has never probed (probing it unauthenticated would fetch live second-stage code; static-read discipline holds).
  • python2 console-script collisions on PyPI/npm as a class (the impersonation-by-script-name trick is copyable).

October 15 one-hundred-and-thirty-sixth sweep: 2.4.2 published ≈17 h BEHIND THE MERGED ADVISORY, ZERO RECORDS, AND THE WHEEL-DIFF IS A C2 TENANCY MIGRATION

pyyi 2.4.2 published 2026-10-14T17:16:43Z, ≈17 h AFTER curated PR #1627 MERGED (00:03:13Z Oct 14) and after OSV MAL-2026-17866 went live, with zero OSV and zero GHSA coverage on 2.4.2 at check (affects=pyyi = 0, 17866 scope still ['2.4.1']). This is the post-advisory out-of-scope iteration class in its sharpest observed form: unlike the @myorder-frontend-commons (112th) or npp-ml cases, the advisory here is neither queue-stale nor scope-drifting, the merged record is CORRECT for 2.4.1, and the code simply moved past it. Advisory-keyed consumers are blind to the live rung no matter which lane writes the record. The infra-scoped hunt (baked space-z.ai slug pair, ptm_ key prefix, /api/cli/* routes) is the only consumer class that covers the iteration.

Wheel-diff (both wheels pulled, unzipped, diffed statically, no execution; 45 changed lines): - DEFAULT_SERVER RE-BAKED from https://preview-chat-960504a9-b4e0-4b69-bc80-6ac3d4fa4ae0.space-z.ai to https://b196q82k8891-d.space-z.ai (FALLBACK_SERVER unchanged, both defaults now converge on b196). This directly answers this page's own monitor question: YES, the baked preview-chat-<uuid> slug rotated with the panel re-deploy. - New functions _homed() + _home_stale_pin(): silent victim re-pinning. Any installed host whose ~/.python2/config.json pins a stale server re-registers against the baked servers and re-pins ONLY if registration succeeds AND the remembered unlock word still unlocks the new home; the stale pin is kept on any failure ("nothing breaks, nothing is asked twice", operator's own comment; new device rows arrive vip; pushes to a stale pin "land in a database the operator's answering surface can never see"). That is a C2 tenancy-migration tool shipped as a point release: the operator moved panels and needed existing implants to follow invisibly. - VERSION bump; run_payload fetch-exec path byte-for-byte identical (exec(compile(...)) still at cli.py:640); PY2_SERVER env override unchanged. - Artifact: pyyi-2.4.2-py3-none-any.whl 9,498 B, sha256 6248e67aa2d1755c9844f299f71948904eaa92fd4bfea324a4ef3ac938603fec vs 2.4.1 9,099 B 00382817….

Registry state at check (~03:3x UTC Oct 15): both 2.4.1 and 2.4.2 served, NOT yanked, PyPI NOT quarantined ≈41.5 h post-OSV, slower than the npp-ml ≈25 h 20 m bound that already closed on this ledger = pyyi is now the slowest advise-to-action clock measured here. #1627 merged; GHSA mirror still absent (OSV-only since birth, 17866 aliases: None).

Durable addition: a point release can be infrastructure churn, not feature churn, when a malicious package's version bump diffs to server-bake rotation + pin-migration logic, the operator's panel moved and every installed host is being herded to the new one. Diffing consecutive wheels of an advised name is a free read on the C2's lifecycle; do it before the next version lands, while both artifacts are still on the index.

October 15 one-hundred-and-thirty-seventh sweep follow-up: MAL-2026-17880 (seapy-ops) BAKES BOTH RAILS INTO AN OSV, THE CLUSTER KEY IS NOW ADVISORY-GRADE

The curated queue drained at 04:02:07Z (#1636 merged) and seapy-ops got MAL-2026-17880 (npm lane n/a; PyPI, scope 1.0.0-1.1.2+1.2.1+1.2.2). Its database_specific.iocs.domains list names preview-chat-960504a9-b4e0-4b69-bc80-6ac3d4fa4ae0.space-z.ai AND b196q82k8891-d.space-z.ai, the exact pair this page treats as the operator's cluster key, including the b196 slug that this wiki first read off the pyyi 2.4.2 wheel-diff ~6 h earlier (Oct 15 136th sweep). Independent confirmation from a second package: (1) the same operator runs at least three names (py2ops → pyyi → seapy-ops); (2) the seapy-ops advisory text independently describes the same grammar family (unlock-gated collection/upload; a REPL mode that ships single-token input to the unlock endpoint and executes a fetched Base64/gzip payload IN MEMORY, disguised as FileNotFoundError); (3) hunt-string value: the two slugs are now queryable inside OSV IoCs, not just off this wiki's page.

Two fleet-relevant reads at check: seapy-ops was advised and NOT quarantined (project-status active, JSON 200, latest 1.2.2 serving), the pyyi slow-class non-action reproduces on the operator's third name, supporting the lane-level-coverage read over per-name oversight; and the advisory's scope misses live rung 1.2.0 (present on the simple index, absent from the enumerated list) = the post-/pre-advisory out-of-scope rung class on this operator too. GHSA mirror still absent at ≈2 h post-birth (affects=seapy-ops&ecosystem=pip = []), until it mints, GHSA-keyed quarantine automation (the ledger's established quarantine trigger) sees nothing. Full sweep record: 137th stream section.

October 15 one-hundred-and-thirty-eighth sweep follow-up: The Lane-Level Non-Enforcement Read Gets Its Mechanistic Proof, The 06:30:33Z Mirror Wave Fired ZERO Seizures

seapy-ops's GHSA mirror GHSA-5j6r-jv3q-c93f minted 2026-10-15T06:30:33Z (same second as the util-foriepft + @flexmr/ui mirrors, the bot batches curated imports), and GHSA-5j6r inherits the OSV's scope gap verbatim: eight rungs 1.0.0-1.1.2+1.2.1+1.2.2, live 1.2.0 absent on BOTH advisory surfaces. One hour after the mint: seapy-ops PyPI NOT quarantined (project-status absent, latest 1.2.2 serving, simple index 15 links), util-foriepft serving with no security-holder recreate, pyyi still latest 2.4.2 unquarantined ≈45 h post-OSV. The ledger's established GHSA-mint→seizure interlock (+57 s/+78 s/+123 s on the August names and the Oct-14 costume) demonstrably did NOT fire on curated-born names. This page's slowest-advise-to-action clock is therefore not an oversight story: the automation that seizes machine-lane names simply has no path to this operator's names, py2ops/pyyi/seapy-ops sit in the lane where neither alias-join consumers nor mint-interlock automation can ever see them. Enforcement here requires a human/abuse action, which makes the infra-scoped hunt (baked space-z.ai slug pair) the only reliable defense for installed hosts. Full sweep record: 138th stream section.