Changelog

#2.10.0 (2026-11-07)

#Changed

• Scans use the full 7,700+ site catalog by default. aliens_eye <username>, the library's scan()/load_sites(), the MCP server, the TUI and --expand now read the extended catalog; --catalog standard (or "catalog": "standard" in config.json) scans the original 788-site list. profile_url() and match_profile_url() still answer with standard-catalog names first, so existing site names keep their URLs. Evaluation and training keep reading the standard catalog, which the ground-truth splits are keyed to.

#2.9.1 (2026-11-07)

#Fixed

• A rate-limit or block page no longer reports an account as missing. A site answering 401, 403, 407, 408, 419, 420, 425, 429, 451, 499, 999 or 5xx serves an error page, not its "no such user" page, yet the detector could still score it Not Found: Hacker News answered HTTP 419 for every username and a real account was reported missing. Those responses are now Maybe (http_unavailable) in scans, site-rule checks and the self-check; 404/410 still mean Not Found. Such pages are also no longer stored as "not found" fingerprints.

#2.9.0 (2026-11-07)

#Changed

• Every catalog site was checked live, with a known account and a made-up one. 945 sites were found broken and are no longer scanned: 68 reported a made-up username as existing (false positives), and 877 failed twice, once on a second pass with longer timeouts (605 unreachable, 272 no longer finding a known account). They are disabled in --catalog extended, now 7,710 enabled sites, and removed from the default catalog, now 788 sites. 56 broken default sites stay because the ground-truth splits use them.
• The detection model is unchanged. Candidates trained on the newly captured pages either raised false positives (7 to 20 on today's holdout pages) or lost recall (0.48 to 0.37), so none beat the shipped model on the holdout.

#Added

• A second pinned community site database (469 valid entries, MIT) is merged into --catalog extended. Same-URL entries merge with the other name kept as an alias, and an active rule replaces an entry its source marks disabled. sites import --include-examples keeps upstream usernames as unverified training examples.
• scripts/train_expanded.py trains a candidate model on expanded-catalog sites: prepare (offline ground truth from upstream and curated usernames), record (live corpus capture), collect, fit and compare (offline, against the holdout). API-probe, disabled, NSFW and holdout-overlapping sites are excluded, and the shipped model is never replaced automatically.
• scripts/verify_catalog.py checks every enabled catalog site live with one known and one made-up username, then judges the frozen responses offline with the scanner's own per-site check, writing a verdict per site to sites.csv. Interrupted captures resume, and apply disables the broken sites (false positives at once, failures only after a recheck).
• train collect --corpus accepts a custom ground-truth file from library callers, and custom ground truth now also excludes holdout host groups.

#2.8.0 (2026-11-07)

#Added

• --explain shows the evidence behind each verdict: the HTTP response, each detector's verdict and threshold, the control check and any browser fallback. HTML and Markdown reports carry the same evidence, and JSON, library API and MCP results gain a structured explanation object.
• --catalog extended scans an expanded community catalog (9,268 entries, 8,587 enabled) built from MIT-licensed sources with recorded provenance, guarded probes and explicit control evidence. The default catalog and model are unchanged. aliens_eye sites import|stats|list|validate inspect and convert local site databases offline.
• --watch-profile name,bio,avatar adds opt-in alerts when extracted profile fields change, comparing only fields with matching provenance.
• CONTRIBUTING.md and CODE_OF_CONDUCT.md.

#Fixed

• Watch mode no longer reports a profile as gone after a timeout, block, error or filter. Absence now needs consecutive HTTP 404/410 responses (--watch-confirmations, default 2), the count survives restarts through scoped state files, and control checks refresh every cycle.
• Diff reports label removals "no longer found" and show profile field changes; malformed nested reports are rejected cleanly.
• --watch with --resume is rejected, --watch --json-stdout emits one report line per username per cycle, and webhooks check the response status and refuse redirects.
• ML validation keeps site aliases and preprocessing from leaking between training and validation folds.

#2.7.5 (2026-10-25)

#Changed

• The note about lxask under the startup banner is now a bright gold box with a star, and the link has its own line so a narrow terminal never splits it. A console that cannot show emoji (an ASCII-only pipe, an old Windows console) gets an asterisk instead of the star. ALIENS_EYE_NO_PROMO=1 still hides it, and it is still never printed with --json-stdout, by the library API or by the MCP server.

#2.7.4 (2026-10-20)

#Fixed

• A scan could run the process out of file descriptors (#26, the follow-up log: Task exception was never retrieved ... OSError: [Errno 24] Too many open files). Three things added up:
  • aiohttp holds each finished request's socket in a pool for 75 seconds (its default), one per site, while the connector's limit counts only the connections in use. Scanning 854 sites therefore left several hundred idle sockets open at once -- past macOS's default of 256 file descriptors, after which every socket, and every attempt to spawn the Playwright driver, failed with Errno 24. Idle sockets are now reaped after 10 seconds, and the CLI raises the soft limit to 4096 at startup where the OS allows it.
  • Errno 24 in a fetch read only "Too many open files", like the site had failed. It now says the process hit its file-descriptor limit and names ulimit -n.
  • A browser fallback that could not start was logged at debug level, so the scan looked healthy while none of its "Maybe" results were re-checked in a browser. The first failure is now warned about once per scan.
• A start of Playwright interrupted by the scan's per-site deadline left its driver process and pipes running, held by nothing: leaked file descriptors, and the fire-and-forgotten Connection.run() task that printed "Task exception was never retrieved". A start that fails or is cancelled now stops the driver it already spawned.

#2.7.3 (2026-10-20)

#Fixed

• --json-stdout printed nothing. It silenced the banner but a scan never wrote its JSON, so aliens_eye username --json-stdout | jq got an empty stream. It now writes the report to stdout as one JSON line per username (JSON Lines), so a single username is one document that json.load or jq reads directly. The banner, progress bar and summary go to stderr, and the output is ASCII-escaped so a console in a legacy code page cannot choke on a name with accents. Files are still written to the output folder as before.
• --only-found did nothing. The flag was read and then ignored. Reports now list just the Found and Maybe sites; the counts in them (sites_scanned, not_found, ...) still cover every site scanned, and the report is marked "only_found": true. It applies to the JSON on stdout and to the saved files, including CSV.

#Added

• A README section and one line under the startup banner about the author's other project, lxask. ALIENS_EYE_NO_PROMO=1 hides the line. It is never printed with --json-stdout, by the library API or by the MCP server.

#2.7.2 (2026-10-19)

#Fixed

• The browser fallback could start dozens of Chromium processes and keep pages open forever (#26, a Mac that ran out of memory with the fallback switched on). The fallback (--playwright, or "fallback" in the interactive settings) re-checks every "Maybe" result in a real browser. Two bugs in it:

  • The browser was started on first use, and the check for "already started" ran before the lock that guards it. With 50 scan workers asking at once, each one launched its own Chromium, and only the last was ever closed.
  • A page that timed out, failed to load, or was cancelled by the scan's deadline was never closed, so a scan accumulated hundreds of open pages.

  The browser now starts once, every page is closed whatever happens, the browser context is replaced every 40 pages so its memory is returned, and a browser that fails halfway through starting is shut down. Checked against a real Chromium: 20 simultaneous fetches, half to a page that never finishes loading, now use one browser (4 processes), 341 MB at most, and leave no page open and no process running after the scan. The old code, given only 8 simultaneous fetches, passed 1.5 GB in seconds.

  Scans without the fallback were never affected.

#2.7.1 (2026-10-19)

#Added

• The Advanced level now says what it will do before it starts. It checks 123 variations of the username, which on all sites is about 105,000 requests and close to two hours. Picking it from the menu now asks for confirmation first (the default is No, and declining returns to the menu). A level given as -l advanced is not asked about. Every level with more than one variation also prints its request count and a rough time under the scan banner. The time assumes about 16 sites a second, measured on a default scan. Prompted by a report of a very long scan on a small laptop (#26).

  Memory, for the record, measured on 2.7.0: a default scan peaks at about 350 MB, and 12 variations of all 854 sites at about 415 MB.

#2.7.0 (2026-10-08)

#Fixed

• Usernames that don't exist were reported as Found (#22, reported with rendered-page evidence by @vollegrewar). A made-up username came back Found on 11 of 19 sites tried, up to 95% confidence. The detector scores how profile-like a page looks, and three kinds of page look the part without being one. A Found now has to survive a check for each:

  • The site's own answer. HTTP 404 or 410 is reported as Not Found whatever the score (chatujme.cz answered 404 and was Found at 84%). Any other error code, such as a 403 block page or a 500, is Maybe at most.
  • The same page for everyone. JavaScript apps, soft "no such user" pages and redirects to search answer every profile address alike. For each Found, the site is now also asked for a username that cannot exist; when both answers are the same page, the result is Maybe. One extra request per Found, cached per site, so a full scan takes the same time (54 seconds before and after for 858 sites).
  • Bot-check pages, as since 2.5.1.

  Measured on recorded pages from 135 sites the model never trained on, same rows before and after:

  	2.6.0	2.7.0	change (95% interval)
  Missing usernames reported Found	33 of 399 (8.3%)	4 of 399 (1.0%)	-7.3 points (-11.7 to -3.5)
  Real accounts reported Found	93 of 172 (54%)	85 of 172 (49%)	-4.7 points (-8.8 to -1.2)
  Precision	0.74	0.96

  The accounts no longer reported Found are ones whose page is identical to a missing user's: nothing in the response told them apart, so they are a Maybe. Of the 4 false hits left, 2 are real accounts: test names meant not to exist that do, on chess.com and Discogs. The other 2 are one site whose pages differ slightly on every load. The threshold was chosen on the 286 training sites, where the same change takes false hits from 7.3% to 0.9%.

  Live, a scan of all sites for a username used on one of them went from 54 Found to 3.

#Added

• --no-control-check (and control_check in a config file) skips the second request.
• aliens_eye selfcheck now scores what a scan reports, with the checks above; --detector-only scores the detector's verdict alone, as before.

#Removed

• Four sites whose address can never identify an account by username: pubmed (a search page), pornmd (a search aggregator), literotica and empornium (the address takes a numeric id). 854 sites remain.

#Not changed

• The other suggestions in #22: rendering every page in a browser, and searching sites that have no profile address. wikipedia and giphy keep their addresses: with the check above, a username that doesn't exist is no longer Found on either.

#2.6.0 (2026-10-08)

#Fixed

• One site could stall a whole scan. A site answering 429 or 503 with a long Retry-After was obeyed in full, so Retry-After: 3600 held the scan at that request for an hour (and the wait was applied twice). The header is the server's to set, so it is now honoured only up to the retry backoff cap (8 seconds); a site asking for longer is reported as an error and not asked again.
• Every site's check now has an overall deadline (about 80 seconds with default settings), after which it is reported as Timeout. A scan is only as fast as its slowest site, and a check that never returned kept the whole scan open.
• YouTube channels were mostly reported as not found. The check asked for youtube.com/<name>, an address only older channels have. It now asks for youtube.com/@<name>, the handle every channel has. Of six real channels that came back "Not Found", five are now found; the sixth no longer exists.
• Cancelling a scan from code that uses the API (for example with a timeout) left its worker tasks running.

#Security

• The address check on scraped URLs could be bypassed with a redirect. Since 2.4.0, avatar URLs taken from scanned pages were only fetched if they pointed at a public address. The check covered the URL itself, and the HTTP client then followed redirects unchecked, so a public URL answering 302 Location: http://127.0.0.1/... was still followed. This affected --correlate and --format pdf. Redirects are now followed one hop at a time with every hop validated, and --correlate pins the connection to the address that was checked, which also closes a DNS-rebinding race.

#Added

• aliens_eye.api.match_profile_url(url): which site and username a profile link belongs to, for any site in the catalogue. https://www.instagram.com/_someone_/ gives ("instagram", "_someone_"); a repository, a navigation page or an unknown host gives nothing.
• aliens_eye.api.fetch_page(url): fetch one page whose address came from somewhere you do not control, with the checks above.
• aliens_eye.api.profile_url(site, username).

#Changed

• --recurse-depth now understands profile links for every site in the catalogue. It used to recognise 19 hard-coded domains and take the first path segment, which missed linkedin.com/in/<user> and <user>.tumblr.com and misread github.com/<user>/<repo>.

#2.5.1 (2026-10-07)

#Fixed

• Fresh installs crashed on start. selectolax 1.0.0 (released 2026-10-03) removed the HTML parser backend Aliens Eye imported, and the dependency had no upper bound, so every new pip install aliens-eye since then failed with ImportError: Modest backend is deprecated since selectolax 1.0. The analyzer now uses selectolax's Lexbor backend, which works on both old and new versions, and the dependency is capped at <2. Checked on 706 recorded pages: no detection result changes. Thanks to @cristianchiriac (#23).

  If you hit this, upgrade: pip install -U aliens-eye.

• Bot-check and error pages reported as Found. Pages such as Cloudflare's "Checking your browser" return HTTP 200 and mention users or profiles, so they could be scored as a found account. A result whose title or meta tags mark it as a challenge, verification or error page is now capped at Maybe. Thanks to @BBmmo-code (#19).

#Added

• 18 Chinese platforms whose profile URL carries the username, including CSDN, SegmentFault, cnblogs, V2EX, LOFTER and Jianshu. Thanks to @BBmmo-code (#19). Two more from that contribution, leetcode.cn and hupu, were left out because they answer "found" for usernames that do not exist.

#2.5.0 (2026-09-18)

#Security

• --format pdf fetched attacker-chosen URLs, including file://. The PDF report downloads each profile's avatar with urllib, which also opens local files. The avatar URL is scraped from the target's own page, so a hostile og:image could point the report at a local file or an internal address. The same public-address check that guards --correlate now guards this path too.

#Fixed

• Every page was parsed from a random-length prefix. The HTTP reader called StreamReader.read(n), which returns up to n bytes -- whatever has arrived on the socket -- not n bytes. Fetching one GitHub profile three times yielded 100000, 100000 and 19497 bytes. On short reads the page's og: tags (at ~21 KB) were missing, so display name, bio and avatar silently vanished, and every detection feature was computed from a different slice of the page on each run. Pages are now read to the size cap or to the end. Avatar downloads in --correlate had the same bug, so images were hashed from partial data.
• Detection model retrained on full pages. The old model had been trained on those truncated prefixes; fed complete pages it got worse (holdout F1 0.543 -> 0.521, false-positive rate 25% -> 29%). The new model is trained on 1,455 samples from 286 platforms -- four times the old training set -- with the blend re-tuned to 0.9 ML / 0.1 heuristic. On 142 platforms it never saw, false "Found" results drop by two thirds (false-positive rate 25% -> 9%, precision 0.46 -> 0.65), at the cost of recall (0.67 -> 0.51). F1 is unchanged within noise. Expect noticeably fewer false leads, and some real accounts moving to Maybe or Not Found.
• --correlate merged almost everything into one "same person" cluster. Correlation linked two profiles when one bio @-mentioned the handle the other was found under. In a scan of a single handle, every profile was found under that handle, and many sites echo it in page chrome (Twitter: "The latest posts from @handle"). One such page therefore linked itself to every other profile: a live scan produced a 228-profile cluster, 227 of whose links ran through Twitter alone. Mentions of a searched handle no longer count. Two bios mentioning the same other handle still link, as does a mention across two different searched handles.
• Display names that are just the site's brand (og:title "Flickr" on flickr.albums) no longer link sibling pages of the same service.

#Corrections

• Results measured before this release, on truncated pages, should be read with that in mind. An earlier claim that ~45% of pages changed within hours compared one random truncation with another and is withdrawn.

#Added

• aliens_eye.api.correlate() also returns profiles: every profile considered, with its avatar hash, so callers can apply their own linkage rules on top of the clusters.

#2.4.0 (2026-09-18)

#Security

• --correlate fetched attacker-chosen URLs (SSRF). Avatar URLs are scraped from the target's own page (og:image, JSON-LD, favicon, per-site selectors), so whoever controls that page chooses them, and they were requested with no validation. A hostile profile could point its avatar at a cloud metadata endpoint, an intranet host, or a service on the analyst's own loopback. Avatar URLs are now limited to http/https and must resolve to public addresses. Upgrade if you use --correlate.

#Added

• aliens_eye.api, a stable programmatic API (scan, correlate, load_sites) for building other tools on Aliens Eye. It returns plain dicts shaped like the JSON report, prints nothing by default, and never writes into the caller's working directory.
• UsernameScanner accepts an on_result(username, result) progress hook (sync or async) and an injectable console.

#Fixed

• Sites whose response headers exceed 8190 bytes (e.g. trakt.tv) failed on every request with Got more than 8190 bytes when reading. The limit is now 64 KB.
• --watch nan and --watch inf were accepted: NaN slips past a <= 0 check because every comparison with it is false. Durations must now be positive and finite.

#Changed

• The MCP server now uses aliens_eye.api, and routes console output to stderr for the whole process, not just during scans.

#2.3.0 (2026-09-06)

Evaluation and reproducibility release. No change to detection behaviour: the shipped model, blend weights and thresholds are untouched.

#Added

• Frozen response corpus (aliens_eye corpus record / corpus stats). Captures raw responses once and replays them through the same fetch path a scan uses, so detector evaluation is reproducible — the same corpus yields byte-identical results regardless of what platforms do afterwards. Bodies are content-addressed and deduplicated.
• Ablation harness (aliens_eye eval ablate). Scores naive status baselines, each judge in isolation, the shipped blend, and per-feature-group ablations over the same corpus rows. Reports precision/recall/F1/FPR plus the Maybe rate, with percentile-bootstrap confidence intervals and paired-bootstrap deltas.
• External tool baselines (aliens_eye eval external). Runs Sherlock's, Maigret's and WhatsMyName's own per-site rules against the same stored responses. Rule data is not vendored; pass it by path and mind the upstream licence.
• Ground-truth builder (aliens_eye eval groundtruth). Expands the ground-truth set from those projects' account lists, recording per-site provenance so no tool is ever scored on the accounts its own rules were tuned against.
• selfcheck --split train|holdout|all, --corpus, and --ground-truth.
• FingerprintStore(read_only=True), required for deterministic replay.

#Changed

• Ground truth expanded from 43 to 428 sites (572 accounts), split site-disjoint into train (286 sites) and holdout (142).
• Corpus captures flush incrementally and support resume, so a long run cannot lose everything on failure.

#Fixed

• Train/eval leakage. Training data and evaluation both derived from data/selfcheck.json, so reported accuracy was a fit measure presented as a generalization measure. Splits are now site-disjoint — by site rather than by account, because the detector learns per-site page structure — and train collect refuses a non-train split without --allow-leakage.
• Documented blend weights did not match the shipped model. WORKING.md stated 0.4*ml + 0.6*heuristic with thresholds 0.6/0.35, while the shipped model.json carries ml_weight 0.6 and thresholds 0.5559/0.3224 and overrides the module constants at runtime. Docs corrected and a test now pins them together.

#2.2.3 (2026-09-06)

#Fixed

• Scans failed outright on many platforms due to a Brotli decoding error. The default request headers advertised Accept-Encoding: gzip, deflate, br, but Brotli was not an install dependency, so any server that honoured br returned a response the HTTP client could not decode (can not decode content-encoding: brotli (br)). Affected sites were reported as fetch errors rather than checked — a silent, total failure rather than a degraded result.

  Measured across a 428-site run, this accounted for 852 of 978 errors, with 173 sites failing on every request, including artstation, bitbucket, anilist, about.me, archiveofourown, allmylinks, behance and bandcamp. All return normal results now.

  Brotli is now a required dependency, and Accept-Encoding is built from the codecs that are actually importable, so an environment predating this release degrades to gzip, deflate instead of failing.

#2.0.0 (2026-06-11)

#Added

• PyPI package: pip install aliens-eye with aliens_eye console command
• Real ML detection: logistic-regression model trained on labeled scans, blended with the heuristic engine; pure-python inference with zero ML runtime dependencies
• aliens_eye train collect / aliens_eye train fit for building datasets and retraining the model
• aliens_eye selfcheck: validates detection accuracy against accounts known to exist
• Rich terminal UI: live progress bar, sorted result tables, summary panels; --plain for scripts
• Proxy support: --proxy (HTTP/SOCKS4/SOCKS5) and --tor
• Site filtering: --site, --exclude-site, --no-nsfw
• Markdown report export (--format md)
• --profile quick|full|aggressive for non-interactive preset selection
• --version, --sites (custom site list), --model (custom model), --no-ml
• Dockerfile
• Test suite (pytest) and GitHub Actions CI (lint + tests on Linux/Windows, Python 3.10–3.13)
• Automated PyPI release workflow via trusted publishing

#Changed

• Restructured to a src/aliens_eye/ package; sites.json and the model ship as package data
• Playwright is now an optional extra: pip install aliens-eye[browser]
• Fingerprint cache and config moved to platform-standard directories (via platformdirs)
• Confidence is now derived from a calibrated probability instead of fixed score bands

#Removed

• Hardcoded /etc, /usr/local, and Termux path probing
• Dead mumble:// site entry

#Compatibility

• python aliens_eye.py from a source checkout still works