Published vulnerabilities
NVD, the CVE List and EUVD supply identifiers, descriptions, CPE configurations, severity and CNA product statements.
Type a product the way people actually write it: Apache 2.4.49, a line of scanner output, a CPE, a package URL. Cull works out which product you mean, which of its advisories apply to exactly that version, and why. We measured both halves against the tools people use today.
Plain product names in, catalogue products out · measured 2026-10-24
To compare the matching itself, we took identification off the table and gave every tool the same 50 exact CPE queries: 25 vendor-documented fixes, each queried once on the last affected release and once on the first fixed one. Exactly the place where a range that is off by one release produces a false alarm or a silent miss. Findings are checked against a frozen reference set; vendor and CNA records take precedence over imported NVD data.
Cull, Search Vulns, cve-search and NVD completed all 50 queries. cve-search ran locally and ignores the submitted version, so it appears here as a version-blind candidate-retrieval baseline: what you get when a product lookup is treated as a version answer. Vulners returned 31 answers before its quota failed.
Extended 2026-08-29 · 25 affected/fixed pairs · exact product CPEs
LOCALHOST identifies the local cve-search run. 31/50 identifies Vulners' partial run.| Tool | Direct query | Practical input workflow |
|---|---|---|
| Cull | Yes | Free text such as Apache 2.4.49, scanner banners, aliases, CPEs, PURLs and package coordinates are accepted directly. |
| Search Vulns | Yes | Accepts a product query directly; this quality run still used the same exact CPE as the other version-aware tools. |
| cve-search | Partial | Requires vendor and product fields and does not evaluate the submitted version. |
| Vulners | No | The measured software-audit endpoint requires an exact software coordinate such as a CPE. |
| NVD | Two steps | Search the separate CPE dictionary, choose the correct CPE, then submit it to the CVE API. The benchmark supplied that CPE. |
50 exact product/version CPEs · Vulners returned 31
Share of known relevant findings returned
Milliseconds · quota pacing excluded
Vulnerability data looks simple until you try to answer one question with it: is this exact thing vulnerable? The answer hides in a dozen feeds that name the same product five different ways, number releases by different rules, and regularly contradict each other. This page walks through what Cull does about that, with the traps we fell into on the way. Most of them you would never think of until a wrong answer showed you.
In short: collect the records, keep what each source actually said, work out which product the query means, compare the version by the rules of the ecosystem it came from, and return the evidence with the answer. Never turn "we could not tell" into "you are fine".
No single feed knows everything, and they disagree more often than you would hope. NVD describes upstream ranges, a Linux distribution knows which fixes it backported, Metasploit knows what actually fires. Cull reads 25 sources and keeps track of what each one is entitled to say.
NVD, the CVE List and EUVD supply identifiers, descriptions, CPE configurations, severity and CNA product statements.
curl, Red Hat, Debian, Ubuntu, Alpine, Wolfi, Chainguard, AlmaLinux and Oracle Linux provide package fixes and backport status; SUSE arrives through OSV.
OSV and GitHub connect vulnerabilities to package URLs and native version ranges for npm, PyPI, Go, crates, Maven, RubyGems, NuGet, GitHub Actions and other registries. Bitnami adds upstream ranges for the server applications it packages.
EPSS, CISA KEV, SSVC, Exploit-DB, Metasploit, Nuclei, PoC-in-GitHub and Wordfence add likelihood, observed exploitation and available tooling.
endoflife.date adds upstream active-support and end-of-life dates for exact product versions. Cull keeps this operational context separate from CVE applicability.
Compare 7.88.1 with upstream's "fixed in 8.x" and you report holes Debian closed months ago. 7.88.1-11+deb12u10 carries the backports. When the query names Debian or the version gives the package away, Debian's own statement decides.
The module tells you exploitation is cheap. Whether your package is affected is still decided by the version evidence. Both facts stay on the finding, neither overwrites the other.
A release upstream abandoned can be fully supported in an enterprise distribution for years. Cull never substitutes one date for the other.
A NetScaler flaw that only applies to appliances configured for SAML is shown with that prerequisite and its Citrix source, and with the honest note that Cull has not checked your configuration.
Native identifiers, aliases, range types, fix statements, withdrawn status, scores, references and exploit provenance stay attached to their source. They are still available when Cull produces a verdict and when someone reviews it.
Every feed uses a different schema, and the dangerous part of an import is not the file that fails to parse. It is the file that parses fine and quietly means something else. The importers translate each schema into Cull's model while keeping the source, the exact range boundaries and the difference between affected, fixed, not affected, withdrawn and unknown.
Under "fixed in 1.4.3" it is: a release candidate sorts before the release. Under "last affected 1.4.2" it is not. Flatten both into "before 1.4.3", as many importers do, and one of them is wrong. Cull keeps them apart.
A format change that makes a parser skip half the records does not crash anything. It just publishes half as many vulnerabilities, which looks exactly like good news. Volume bounds and skip ratios make that build fail instead.
OSV sometimes bounds a range with git commits. Cull cannot compare a release number with a hash, so it marks the statement unverifiable instead of guessing, and the finding says why it could not decide.
NVD files it as version 1.0.2 with update u. Join them with a dash and 1.0.2-u becomes a pre-release that sorts below 1.0.2. Cull attaches a single letter directly, so the patch level sorts above the release it patches.
This is where most vulnerability tools quietly lose findings. The same software appears as an NVD CPE, an SBOM package URL, a distribution package, an EUVD label and two former company names. Every one of those keys holds part of the product's history. Search one of them and the advisories filed under the others are simply invisible, and the result looks clean.
NVD files projects without a company under check_mk_project and later under the company's own name checkmk. Varnish once reported not affected for 6.0.0 from the three-record spelling while Varnish Cache had eleven advisories for that release.
Sun's old sun:virtualbox entry holds five advisories; Oracle VM VirtualBox holds 245 for that release. Ask the old name and you get an all-clear. Lotus Domino lives on as ibm:lotus_domino, JBoss as Red Hat's.
Cull therefore groups identifiers into one product entity while building the dataset. The links come from ten places, each with its own evidence:
sd_675_firmware and sd675_firmware are linked. A changed letter or meaningful symbol is not ignored: 865 and 865+ are different chips.palo_alto_networks and paloaltonetworks, gpac and gpac_project, kaspersky and kaspersky_lab: words that only describe the organisation do not make another vendor. The product itself still has to match.lotus:domino is ibm:lotus_domino. A generic two-word name that many vendors carry, such as web_server, never qualifies.deprecatedBy field when it describes a clean one-to-one rename. A key that splits into several successors is left alone.npmjs.com/package/lodash, that is direct evidence that it names the npm package.iOS for NVD's iphone_os, k8s for Kubernetes, MSSQL for SQL Server.It shares the name and some CVEs, but numbers its releases 0.x while the Java line is at 1.11. Merge them and every Rust release becomes "affected" by every Java bug below 1.11. Requiring the same fixed releases keeps bindings, bundled copies and PubNub's per-language SDKs apart.
The WordPress theme ask-me carries the CVEs of two different vendors' "Ask Me" products. Linking both through it would quietly merge them, so it links nothing. Reviewed controls such as curl versus libcurl, Firefox versus Firefox ESR and FileZilla client versus server must stay apart, or the build does not ship.
A union-find pass merges the accepted links so searches do not have to walk the graph. No product may collect more than 64 identifiers. If a bad rule crosses that limit, the build fails instead of publishing an oversized product group.
Bundles and dependencies often appear in the same advisory: Chrome and the Skia library it contains share CVEs and stay separate. Shared CVEs only count together with a matching name and matching fixed releases, and reviewed exclusions keep known mixed EUVD keys apart.
People do not type CPEs. They paste OpenSSH 7.9p1 Debian 10+deb10u2 from nmap, microsoft-iis-version:Microsoft-IIS 10.0 from a scanner template, or just Unity 2022.1.25f1. Free text is ranked with BM25; exact CPEs (2.3 and the older cpe:/ form), package URLs, distribution package versions and advisory ids such as CVE, GHSA, EUVD, DSA or USN skip that step. Bulk search accepts up to 100 targets.
"Unity" is a game engine, Dell's storage arrays, a Cisco voicemail system and a Linux desktop shell. Cull answers with the most likely product and weighs real evidence to pick it: software beats a hardware-only entry, a product whose advisories number releases on the query's scale beats one that does not, then endoflife.date's curated name, a vendor written in the query, and finally a twentyfold lead in advisories. Unity 2022.1.25f1 is not an array whose firmware is at 5.x.
Unless that evidence is decisive, the answer is flagged product_uncertain, the other products stay one click away, and the result can never become an all-clear. That rule exists because favicon once resolved to an obscure plugin of that name and reported "nothing known", as if the reader's software were clean.
The "9" is BIND's release line, and the entry named "BIND 9" holds a fraction of the advisories. Cull reads it like a version branch and answers with BIND itself. Windows 10 stays Windows 10, because that is a product in its own right.
nmap prints the Debian revision as separate words. Cull joins them back into 7.9p1-10+deb10u2, recognises Debian 10 from deb10u2, and judges the version against Debian's backports instead of upstream's ranges. Exim smtpd and Squid http proxy lose the service words the same way.
Neither is Oracle WebLogic Server Oracle Linux. Both words count as a distribution only with a release after them. And adobe is a real npm package, which is exactly why a bare vendor name never silently becomes it.
NVD filed one Liferay Portal advisory under liferay:liferay and 323 under liferay_portal. An exact name with almost nothing behind it gives way to the vendor's name plus a classifier: Liferay Portal, Cisco RV320 Firmware. "KDE" does not become Konqueror; it stays, marked uncertain.
Putting every alias into one BM25 document penalizes products with many names. That once caused apache to fall out of the top results after httpd was linked to it, and joining a name's spellings into one document made node.js rank on the term "js" alone and land on axios. Cull now stores the best score for each individual name when it builds the index. Adding an alias cannot push the product down.
A search for postgressql returns no findings and offers PostgreSQL as a correction. It does not silently run the corrected query: someone handed a vulnerability list for a product they did not ask about has no way to notice, because the CVEs are real. Short inputs receive no correction because one changed character can name another real product; curl and cull are a good example. A partial match is only answered when its name opens the query, so OpenAI API Models is not the npm package api-models.
Query parsing also has to distinguish names from versions. Windows 10 1809 means Windows 10 at build 1809, not Windows at version 10. Distribution names are removed before version detection, so the final 12 in Squid 5.7-2 Debian 12 belongs to Debian.
| Constant | Value | Why |
|---|---|---|
| k1 | 0.9 | Below the usual 1.2. Product names are short, so a repeated word says little. |
| b | 0.6 | Length normalization, kept mild so an entity with many aliases is not penalized. |
| exact-name | 2.5x | A whole name equal to the query outranks a mere term match. |
| prior | up to 1.6x | Advisory volume breaks ambiguous textual ties without overriding a strong name match. |
Finding the product is half the job. Cull must then place the submitted version inside or outside every relevant advisory range, and "is 1.2 before 1.3" turns out to have about a dozen answers. It parses the version into segments and chooses comparison rules from the package ecosystem or distribution it came from.
1.2rc1 sorts before 1.2, while 1.2p1 sorts after it. The same string can change meaning between ecosystems: SemVer treats 1.0.0-1 as a pre-release, while Debian treats the suffix as a package revision. Maven, RubyGems, NuGet and Alpine have their own ordering rules as well.
Red Hat ships Apache with epoch 1, written 1:2.4.37, but rpm -q may print only 2.4.37-43.el8_5. Read that as epoch 0 and a fixed server looks vulnerable to everything since 2018. If only one side has an epoch, Cull compares the visible version and release; if both do, normal RPM ordering applies.
So are OpenSSH 9.7p1 and 9.8p1, and sudo 1.9.5p1 and 1.9.5p2. NVD stores the suffix in a separate CPE field; Cull folds it back so the release you type is the release that is compared. All three pairs are in the benchmark for exactly that reason.
Both share one CPE and differ only in its edition field. A statement scoped to Enterprise must not confirm, or clear, a Community query, so the edition is kept as context rather than folded into the product.
apache 2.4.49 does not borrow Debian's fixesDistribution rules apply only when the query names a distribution or contains a recognizable package version such as 7.88.1-10+deb12u5. If the only evidence belongs to some other distribution, Cull leaves the result undecided rather than picking one.
Each relevant advisory receives a result and a reason. The most important design choice on this page: unknown stays separate from ruled out, so a lack of usable range data can never look like a clean result.
In automation an empty result looks like success. So an unknown advisory id is reported as unchecked instead of dropped, a product Cull could not identify is UNRESOLVED rather than "not affected", and an uncertain product match can only ever produce "unknown".
A distribution may overrule an upstream match only with positive evidence: fixed, not affected or out of range. An inconclusive or withdrawn record cannot, and a finding is only withdrawn when every source that reported it has retracted.
The API exposes the same distinction as AFFECTED, NOT_AFFECTED, UNKNOWN_OR_UNCHECKED and UNRESOLVED.
An affected result does not say whether anyone can actually exploit it this afternoon. Cull keeps exploit evidence separate from CVSS and shows where that evidence came from.
EPSS estimates likelihood. CISA KEV records observed exploitation. SSVC provides decision context. Exploit-DB, Metasploit, Nuclei and PoC-in-GitHub indicate what public tooling exists. Wordfence adds coverage for WordPress plugins that often have no CVE at all.
The strongest evidence determines the finding's exploit tier:
Results are ordered by this evidence before severity: known-exploited first, then any exploit, then EPSS probability, and only then CVSS. A 9.8 that nobody can reach sits below a 7.5 with a maintained Metasploit module. Severity, EPSS, publication date and identifier remain available as explicit sort keys, and the tier itself is a property of the finding that does not change with the order. If Cull has a validated Nuclei template id, Metasploit module path or Exploit-DB id, it can show the corresponding command. A reference URL is never inserted into a shell command: a link in an advisory is somebody else's text.
CVSS describes technical severity, EPSS estimates likelihood, KEV records observed exploitation and SSVC adds decision context. If one of those values is missing, Cull leaves it missing rather than filling the gap with another score.
Cull's research workbench keeps the finding list next to the selected CVE. Applicability, source statements, aliases, affected ranges, fixes, exploit tooling and the event timeline stay in one place, so you never need five tabs to answer one question.
Search ids, descriptions, CWE, products, sources and tooling. Structured filters cover CVSS, EPSS, KEV, exploit type, certainty, fixes, dates, vector fields and attack primitives.
The inspector shows source verdicts, disagreements, remediation, affected branches, commands and the timeline for the selected finding.
Save findings locally in the browser, add comments, reorder them and filter the queue by KEV, tooling or severity.
Copy selected findings as Markdown, CSV or JSON. Queue exports include notes, and verified Nuclei templates can be combined into one command.
The query language and the controls use the same fields. openssl 1.0.1 is:kev cvss:>=9 vector:AV:N can be bookmarked and rerun. Filters are applied before totals, sorting and pagination, so the displayed count matches the filtered result.
CVSS, EPSS, known exploitation, exploit maturity, fix availability and publication date remain separate fields. You can choose which one matters for the current job.
The HTTP API calls the same matcher as the web interface. It accepts the same targets and filters, returns the same verdict states and adds pagination and stable response schemas for integrations.
GET /api/v1/searchSearch one product, package coordinate or advisory id. Results can be filtered and sorted, or reduced to ids with view=ids.
POST /api/v1/bulkCheck up to 100 targets in JSON or newline-delimited text. Compact output is available for inventories and boundary checks.
GET /api/v1/openapi.jsonDownload the OpenAPI 3.1 description of parameters, result states, findings, pagination and errors.
The full response includes aliases, source verdicts, affected and ruled-out evidence, unchecked records, CVSS, EPSS, KEV, SSVC, CWE, attack primitives, exploit references, commands, timelines, fixes and source disagreements. CI jobs that only need identifiers can request view=ids.
The status field removes ambiguity around empty results. AFFECTED has applicable evidence, NOT_AFFECTED resolved the target without one, UNKNOWN_OR_UNCHECKED has relevant but inconclusive evidence, and UNRESOLVED could not identify the product. A product chosen without decisive evidence carries product_uncertain, so an inventory tool can ask a human before it stores the mapping.
The workbench, single-search endpoint and bulk endpoint render the same underlying result. Browser and API behavior therefore change together.
Cull serves about 2,172,616 advisories from immutable in-memory structures on one machine, which is why a complete answer arrives in about 18 milliseconds. In the measured 2026-08-24 corpus, the SQLite database was 2.4 GB and the separate prose file 558 MB. The loaded server used about 1.6 GiB of live Go heap and 2.5 GiB steady RSS.
Structured matching data lives in the Go heap. Titles, descriptions and reference URLs live in a separate memory-mapped file because the matcher does not need them while deciding applicability. The file contains no Go pointers, so the garbage collector does not scan it, and the operating system can discard and reload those pages under memory pressure.
Repeated values are interned. In that corpus, 12.8 million range-list occurrences reduced to 311 thousand distinct values, and 3.1 million parsed-version occurrences reduced to 250 thousand. Parsed versions use one allocation each. Temporary build maps are discarded before the server reports ready.
GOMEMLIMIT cannot shrink live data. A value below the working set only makes garbage collection more frequent. The actual savings come from compact structures, interning and the pointer-free mapped file.
Every example on this page was once a wrong answer. Most matching bugs live at boundaries: a rename, an exclusive upper range, a package epoch, a backport, an alias collision, a scanner that splits a version in two. Once one is fixed, it stays in the regression suite, so it cannot come back.
A distribution package named exactly "mysql 8.0" won the version-bearing reading, was then set aside as packaging, and the query was refused although "mysql" matched perfectly. Now every reading is kept and the next one answers.
Both stopped resolving once the catalogue held a source-local entry of exactly that word. The vendor's own product named after itself now wins, while "adobe" still asks.
The release branch in the name found SUSE's parallel package openssl-1_1. A release branch read into a name can no longer pick distribution packaging the query did not ask for.
Read word by word it became Alpine context around a product called "linux". Distribution names are now read as whole phrases, and a full name such as Fedora Linux stays a product.
Hundreds of focused automated tests cover version rules, query parsing, product links, source precedence, verdicts, filters, pagination and API schemas.
Corpus fixtures run through import, matching and response rendering. Tests check the bucket, reason, evidence, aliases, fixes and status around an id.
Freshness, source volume, parse coverage, corpus shrinkage and product fan-out are checked before a dataset can serve searches, and reviewed identity controls must hold.
The 50-query benchmark uses fixed inputs and a reviewed reference set, and thousands of real product names from scanners and catalogues test identification. Both record misses separately from false alarms.