Skip to content

AAK-KNOWN-VULN-001

A locked, pinned or installed release is affected by a published advisory

Field Value
Severity The advisory's: CRITICAL from CVSS 9.0, HIGH from 7.0, MEDIUM below that; INFO for a range
Category SUPPLY_CHAIN
Shipped v0.6.29
Scanner agent_audit_kit/scanners/known_vulns.py
Feed agent_audit_kit/advisory_feed.py, NVD's CNA records through agent_audit_kit/nvd_feed.py
OWASP MCP MCP04:2025 (Software Supply Chain Attacks & Dependency Tampering)
OWASP Agentic ASI04 (Agentic Supply Chain Vulnerabilities)

Why it exists

On 26 and 27 Oct 2026 advisories were published against vLLM (below 0.32.0), MLflow (3.12.0rc0 up to 3.17.0), the OpenHands Software Agent SDK (below 1.52.0) and NVIDIA's NeMo Agent Toolkit. A project that locked one of those releases passed every rule on 0.6.28, because AAK knew a vulnerable release only through a hand-written pin. Meanwhile aak update was already downloading OSV's complete PyPI and npm exports for the malicious-release feed, and discarding every record that was not MAL-. This rule keeps those advisory records and checks every release a scan resolves against them.

Two days later, on 28 and 29 Oct, IBM published CVEs against Langflow OSS 1.0.0 through 1.12.8, scored 9.8. OSV had no package record of them, so the feed knew nothing, and a project that locked langflow 1.12.8 passed every rule on 0.6.29. The one machine-readable range was IBM's own, in NVD's record of each CVE. The feed now also reads those CNA records, for the CVEs OSV has no package record of (see Sources).

What it catches

A PyPI, npm, Go or crates.io release that an advisory lists as affected, wherever the project resolves it:

  • a lockfile that resolves it: uv.lock, poetry.lock, Pipfile.lock, package-lock.json, npm-shrinkwrap.json, pnpm-lock.yaml, yarn.lock (v1), Cargo.lock (a block with a source line: the root crate, path dependencies and workspace members have none and are the project's own code);
  • a go.mod that requires it: every require line, direct or // indirect. A replace to another module and version is judged as that module, one to a local path drops the module, and exclude is ignored. go.sum is not read: it lists modules outside the build list. A go.mod older than go 1.17, or with no go line, lists only part of the build, and a finding on one says so;
  • a declaration that pins it: vllm==0.31.1 in requirements.txt, an exact version in pyproject.toml, Pipfile or package.json, or a launch such as uvx pkg==1.2.3 or npx -y pkg@1.2.3 in an MCP config;
  • a range every release of which is affected, such as vllm<0.32;
  • an installed <name>-<version>.dist-info in a site-packages directory, a project's .venv and an unpacked image root filesystem included;
  • with aak scan --include-user-config, a release a launcher keeps on this host: a package in npm's npx cache (<cache>/_npx/<hash>/node_modules/<name>) or a dist-info in one of uv's tool environments (<cache>/environments-v*/), read as AAK-MALICIOUS-RELEASE-001 reads them. The evidence names the cache path. Without the flag nothing outside the project is read.

A range that admits affected and unaffected releases, such as vllm>=0.30, is one INFO finding naming the advisories it admits, so a CI gate on HIGH does not fail on a range that may resolve to a fixed release. When a lockfile in the same directory resolves the package, the range is not reported at all: the lockfile says what is installed. A declaration without a version is not judged, except an MCP config's launch: one that pins no release (no version, a tag or a range) is AAK-KNOWN-VULN-003.

The finding names the advisory, its CVE ids, its score, the release and the release that fixes it; a sample is under Related advisories. When the fix is on a date line npm has since deprecated, the finding says so, for the packages agent_audit_kit/data/date_line_packages.json names: the four npm reference servers went back to SemVer at 1.0.0 in October 2026, so 2025.7.1, the fix an older advisory names for server-filesystem, is not where to go any more.

The feed

A scan reads ~/.agent-audit-kit/advisories.json when aak update has written it, else the records bundled with AAK, and never touches the network. The SARIF run's properties.dataFeeds.knownVulnerabilities and the console summary say which copy was read and when it was generated.

  • aak update builds the cached copy in the same pass that rebuilds the malicious-release feed, from the same four downloads (PyPI, npm, crates.io and Go, about 280 MB). It keeps every record that is not MAL-, not withdrawn, and has an ECOSYSTEM or SEMVER range or a list of affected versions. On 2026-10-29 that was 23,579 package entries for 5,392 packages (PyPI 15,252, npm 8,327), 8.7 MB.
  • A RustSec record marked informational (unmaintained, unsound or notice, 521 of the 3,055 crates.io records on 2026-11-07) is left out: RustSec files those as notices and cargo-audit reports them as warnings.
  • A cached copy over 60 MB would keep only HIGH and CRITICAL advisories, and aak update and the summary would say so. The feed is well under it today.
  • The bundled copy holds four OSV records exactly as OSV returned them: on 2026-10-29, GHSA-h3rc-j82p-2c3h (vLLM) and GHSA-26p8-2jq9-3vq9 (MLflow), and on 2026-11-07, GHSA-82r6-8w77-94w6 and GHSA-5p39-cfhj-2xmp (anyio below 4.14.2). OSV had no record for the OpenHands SDK or the NeMo Agent Toolkit advisory that day, so neither is bundled; once OSV publishes one, aak update picks it up. It also holds the entries nvd_feed converts from NVD's records as NVD served them on 2026-10-30, 2026-10-31 for lmdeploy and 2026-11-07 for crates.io's surrealdb and Go's mcp-go and ChatCLI (saved unedited under tests/fixtures/nvd_cna, and listed under Related advisories): three against langflow, one against lmdeploy, and one against the npm packages open, @modelcontextprotocol/inspector and opencode-ai. Every one of the five maps to a package, so none was left out. Without aak update, nothing else is known.

Sources

The feed has two sources, and a finding says which one it came from.

OSV, read as above: every PyPI and npm advisory record in OSV's exports.

NVD's records of CNA version data, for the CVEs OSV has no package record of. A CVE's CNA (IBM for Langflow, VulnCheck, GitHub) writes which product versions are affected in its CVE record, and NVD's CVE API 2.0 carries that data at vulnerabilities[].cve.affected[].affectedData[]: a vendor, a product, sometimes a packageURL, a collectionURL and packageName, or a repo, and versions[] of version, lessThan or lessThanOrEqual, versionType and status. NVD's own CPE configurations, added once NVD has analysed a record, are not read: a CPE names a vendor and product, not a package.

  • What aak update reads. Every NVD record modified in the last 120 days, NVD's longest window (aak update --nvd-days N reads fewer; --nvd-days 0 reads none), 2,000 a page, pausing between requests as NVD asks: six seconds without an API key, less with NVD_API_KEY in the environment, which is sent as a header and never printed. On 2026-10-30 a 120-day window held 100,047 records, 51 pages of about 4 MB at about 20 seconds each: plan for 20 minutes and 200 MB. The entries read replace the NVD entries read before, so the feed holds the NVD records modified in that window.
  • Which package a record names, in this order and never from a name alone: the CNA's packageURL when it is pkg:pypi/... or pkg:npm/... (a packageURL of another registry is not overridden); else its collectionURL when that is PyPI or npm: a package's own page (https://pypi.org/project/<name>/, https://www.npmjs.com/package/<name>, with a packageName, if any, naming the same package), or the registry's root as the CVE JSON 5 schema's examples write it (https://pypi.python.org, https://registry.npmjs.org) with the name in packageName. A page and a packageName that disagree map nothing, and the record is counted unmapped rather than guessed at; any other collectionURL (a GitHub URL, a vendor's site) is passed over. JFrog's prefab-ui record (in the table below) names its package this way and no other. Else a row of the reviewed map, agent_audit_kit/data/cna_package_map.json, whose repository is the CNA's repo, or whose vendor and product are the CNA's. A row is added only after checking that the package's own page on pypi.org or npmjs.com names the same repository. On 2026-11-02 the map has four rows: IBM's "Langflow OSS" is PyPI langflow, the repositories sindresorhus/open and modelcontextprotocol/inspector are npm open and @modelcontextprotocol/inspector, and VulDB's "samanhappy MCPHub" is npm @samanhappy/mcphub. A record that maps to no package is counted and dropped, and aak update prints how many and the ten (vendor, product) pairs that most often mapped to nothing. A record for software on neither registry can map to a launch row of the same map instead, which AAK-KNOWN-VULN-002 reads.
  • Which versions. Only status: affected entries, read on the registry's order whatever versionType says of semver, custom or python (a git commit is not read): version with lessThan is affected up to, not including, that release; with lessThanOrEqual, up to and including it, which is how IBM wrote "1.0.0 through 1.12.8". A version with neither is one release when it is one, and otherwise read as GitHub's advisory grammar, the only way GitHub's CNA writes a range: < 0.5.0 is affected below 0.5.0, >= 2.2.0, < 3.0.1 from 2.2.0 below 3.0.1, <= 1.2 up to and including 1.2, >= 1.0.0 from 1.0.0 on, and = 1.4.2 that release alone. > x starts at the first version after x, which semver has (3.0.1-0 after 3.0.0) and PEP 440 does not, so on a PyPI package it is not read. Other text (1.0 to 1.8.0, two clauses with no comma) is not read, and a record left with no range is counted. changes split a range where the status moves. defaultStatus: affected makes every release affected but those the entries exclude, read only when each exclusion is a range ending before a release or open-ended; a record whose exclusions no interval can state is not read rather than reported wider than the CNA wrote it.
  • Severity. The record's CVSS 3.x base score, computed from its vector, NVD's own (Primary) metric before the CNA's; without one, the CVSS 4.0 metric's label; without either, MEDIUM. The finding says "NVD CNA record" and links the record on nvd.nist.gov.
  • OSV wins. An NVD entry is dropped when an OSV entry for the same package names its CVE, as its id or an alias, so a CVE is reported once, with OSV's ranges, as soon as OSV has a record of it.
  • When NVD cannot be read, aak update rebuilds the OSV half, keeps the NVD entries scans read before (the cached feed's, else the bundled ones), says so, and exits non-zero. A scan never reads NVD.

How a release matches

  • In the order its registry uses: PEP 440 for PyPI, so MLflow's >=3.12.0rc0, <3.17.0 takes 3.12.0rc0 and 3.17.0rc1 and leaves 3.17.0, and semver for npm.
  • Against the record's ranges, read in version order as OSV's specification evaluates them, and its listed versions. A git range, which names commits, is not read.
  • A PYSEC record and its GHSA twin are one advisory, reported once under the GHSA id with both ids named. Two GHSA records are never merged, even when a PYSEC record names both.
  • One finding per file, package and advisory, at the first entry that resolves an affected release. A lockfile is read for its first 64 distinct releases of a package.

Severity

The severity is the advisory's, not the rule's: CRITICAL from a CVSS 3.x base score of 9.0, HIGH from 7.0, MEDIUM below that. The score is computed from the record's CVSS 3.x vector. Without one, GitHub's label decides (CRITICAL and HIGH as they are, MODERATE and LOW as MEDIUM), and a record with neither is MEDIUM. A CVSS 4.0 vector is not scored. An NVD entry is scored as Sources says: its CVSS 3.x vector, else its CVSS 4.0 label, else MEDIUM.

The score is the record's own. vLLM's record carries GitHub's vector (AV:N/AC:H/..., 8.1), so vLLM 0.31.1 is HIGH, although NVD scored the same flaw 9.8.

Version pins

A hand-written pin is a judgement someone read the advisory for. Where another rule reports every CVE id of the advisory on the same line of the same file (the declaration or lockfile entry both point at) at or above this finding's severity, only that rule's finding is kept: on 2026-10-29, 82 of the 134 pin-table rows named a CVE that OSV also lists. Elsewhere this rule still reports, for instance in a lockfile below the root, which the pin table does not read, in an installed dist-info, which no pin reads, or where a configuration rule cites the CVE on another line of the file.

How to fix

Upgrade to the release the finding names as fixed on that line, re-lock, and rebuild the environment or image. For a manifest range, raise its lower bound past the fix. Where the advisory lists no fixed release, remove or replace the package, or keep the component the advisory names unreachable until one ships. Run aak update so the scan reads the current advisories rather than the bundled few.

The advisories that showed the gap, as NVD and OSV listed them on 2026-10-29 (the first four), 2026-10-30 (the Langflow and open records) and 2026-10-31 (lmdeploy). They are context, not this rule's references: the rule reports whatever advisory the feed holds.

Product CVE OSV record In the bundled feed
vLLM below 0.32.0 CVE-2026-115915 GHSA-h3rc-j82p-2c3h (CVSS 3.x 8.1) yes
MLflow 3.12.0rc0 up to 3.17.0 CVE-2026-116692 GHSA-26p8-2jq9-3vq9 (CVSS 3.x 9.8) yes
anyio below 4.14.2, TLS host names CVE-2026-63374 GHSA-82r6-8w77-94w6 (GitHub rates it CRITICAL, CVSS 4.0 only) yes
anyio below 4.14.2, process-pool workers CVE-2026-64847 GHSA-5p39-cfhj-2xmp (GitHub rates it MODERATE, CVSS 4.0 only) yes
OpenHands Software Agent SDK below 1.52.0 CVE-2026-112346 none yet no; AAK-AGENT-SERVER-NOAUTH-001 reads its launchers
NVIDIA NeMo Agent Toolkit CVE-2026-65145 none yet no
Langflow OSS 1.0.0 through 1.12.8 CVE-2026-110186 none yes, from IBM's record in NVD (CVSS 3.x 9.8)
Langflow OSS 1.0.0 through 1.12.8 CVE-2026-93680 none yes, from IBM's record in NVD (CVSS 3.x 9.8)
Langflow OSS 1.0.0 through 1.12.3 CVE-2026-104339 none yes, from IBM's record in NVD (CVSS 3.x 9.8)
npm open 7.2.0 up to 11.0.0, @modelcontextprotocol/inspector 2.0.0 up to 2.12.0, opencode-ai 1.0.137 up to 1.18.33 CVE-2026-119025 none yes, from VulnCheck's record in NVD, which names each package by packageURL (CVSS 3.x 7.5)
prefab-ui 0.14.0 up to 0.21.0, the Generative UI sandbox FastMCP's generate_prefab_ui runs CVE-2026-120338 none (GHSA-595q-jv4r-q7j9 is a repository advisory only) yes, from JFrog's record in NVD, which names the package by collectionURL and packageName (CVSS 3.x 10.0)
@samanhappy/mcphub 1.0.0 through 1.0.45 and 1.1.0 through 1.1.7, fixed in 1.1.8 CVE-2026-120323 none yes, from VulDB's record in NVD, which lists each release and maps by the reviewed row (CVSS 3.x 4.3); AAK-MCP-MCPHUB-CVE-2026-79748-001's floor reports the same releases, so its finding is the one kept
Go github.com/mark3labs/mcp-go v0.48.0 up to v1.2.1, a malicious MCP server gets the user's token minted for another resource CVE-2026-123272 none (GHSA-xf6j-693p-wqxx is a repository advisory only) yes, from VulnCheck's record in NVD (CVSS 3.x 8.0), read 2026-11-07
Go github.com/diillson/chatcli v1.159.0 up to v1.215.2, the same flaw in ChatCLI's MCP OAuth login CVE-2026-123271 none (GHSA-95vg-3r9c-2v5q is a repository advisory only) yes, from VulnCheck's record in NVD (CVSS 3.x 8.0), read 2026-11-07
crates.io surrealdb 3.1.0 up to 3.3.0, an MCP session takeover with a low-privilege token CVE-2026-121419 none (GHSA-pgh9-49q4-hqcx is a repository advisory only) yes, from VulnCheck's record in NVD (CVSS 3.x 8.1), read 2026-11-07
InternLM LMDeploy 0.9.0 through 0.18.0, no fixed release CVE-2026-119706 a git commit-range record only, none for PyPI lmdeploy yes, from VulnCheck's record in NVD, which names the package by packageURL (CVSS 3.x 9.8)

A finding on the vLLM one reads:

uv.lock resolves vllm 0.31.1, which GHSA-h3rc-j82p-2c3h (CVE-2026-115915, CVSS 3.x 8.1)
lists as affected; it is fixed in 0.32.0 (https://osv.dev/vulnerability/GHSA-h3rc-j82p-2c3h).

and one on a Langflow record NVD carries:

uv.lock resolves langflow 1.12.8, which CVE-2026-110186 (NVD CNA record, CVSS 3.x 9.8)
lists as affected; the advisory lists no fixed release, and 1.12.8 as the last affected
one (https://nvd.nist.gov/vuln/detail/CVE-2026-110186).

Registry reference

This section is generated from the rule registry (agent_audit_kit/rules/builtin.py) when the documentation is built, so it follows the scanner. The rest of the page is hand-written; where the two disagree, the registry is what runs.

Severity
CRITICAL
Category
Supply Chain (supply-chain)
Scanner
agent_audit_kit/scanners/known_vulns.py
Auto-fixable
No.
SARIF name
KnownVulnerableRelease

What it detects

A lockfile resolves, a manifest pins, or a site-packages directory has installed a PyPI or npm release, or a Cargo.lock resolves a crates.io release, that an advisory lists as affected. vLLM 0.31.1 and MLflow 3.16.0 locked in a project passed every rule on 0.6.28 (28 Oct 2026), although CVEs against both were published on 26 and 27 Oct, because AAK knew a vulnerable release only through a hand-written pin. The rule reads the offline advisory feed aak update builds from the OSV exports it already downloads for the malicious-release feed, plus OSV's crates.io export: every record for those registries that is not MAL-, withdrawn or a RustSec notice and has a range or a version list. OSV can lag a CNA by days: IBM's Langflow OSS CVEs of 29 Oct 2026, CVSS 9.8, had no OSV package record while NVD carried IBM's affected range, so the feed also holds entries converted from NVD's records of CNA version data, for the CVEs OSV has no package record of. SurrealDB's MCP session takeover of 3 Nov 2026 (CVSS 8.1) is one: NVD's record names pkg:cargo/surrealdb, and OSV has none. A scan reads the copy cached at ~/.agent-audit-kit/advisories.json, else the records bundled with AAK, and never touches the network. The finding carries the advisory's severity: CRITICAL from a CVSS 3.x base score of 9.0, HIGH from 7.0, MEDIUM below that or unscored, and INFO for a manifest range that admits affected and unaffected releases.

How to fix (registry)

Upgrade to the release the finding names as fixed on that line, re-lock, and rebuild the environment or image; for a manifest range, raise its lower bound past the fix. Where the finding says the fix is on a date line npm has deprecated, take a current SemVer release instead. Where the advisory lists no fixed release, remove or replace the package, or keep the component the advisory names unreachable until one ships; where an NVD entry names only the last affected release, upgrade past it to the release the vendor's advisory names as fixed. Run aak update so the scan reads the current advisories rather than the bundled few. Where the finding says a parent pins the release exactly, the upgrade needs an override: [tool.uv] override-dependencies or a pip constraint.

To confirm the finding is gone, report only this rule:

agent-audit-kit scan . --rules AAK-KNOWN-VULN-001

Limitations

Reads what AAK-MALICIOUS-RELEASE-001 reads, anywhere outside AAK's skip directories: the lockfiles uv.lock, poetry.lock, Pipfile.lock, package-lock.json, npm-shrinkwrap.json, pnpm-lock.yaml and yarn.lock (v1); and Cargo.lock, for crates.io advisories, which that rule does not read; declarations in requirements*.txt, pyproject.toml, Pipfile, package.json (not its scripts) and the launches in MCP configs; and <name>-<version>.dist-info directories in the site-packages and dist-packages directories AAK-STARTUP-HOOK-PTH-001's walk reaches; with --include-user-config also the packages in this host's npx cache and uv's tool environments, read as AAK-MALICIOUS-RELEASE-001 reads them and reported with the cache path. A release matches by PEP 440 order on PyPI (3.12.0rc0 is inside >=3.12.0rc0, 3.17.0rc1 below a 3.17.0 fix) and semver on npm, Go and crates.io, against the record's ECOSYSTEM and SEMVER ranges, read in version order, and its listed versions; a git range, a CVSS 4.0 vector and a record with no PyPI, npm, Go or crates.io package are not read, and a record scored by neither a CVSS 3.x vector nor GitHub's label is reported MEDIUM. A Cargo.lock block counts only with a source line, so the root crate, path dependencies and workspace members are left out, and a git dependency is judged by its crate version. Cargo.toml is not read, so a crate with no Cargo.lock beside it is not judged. A RustSec record marked informational (unmaintained, unsound or notice, 521 of 3,055 on 7 Nov 2026) is left out, so an unsound crate with no GHSA record is not reported. A go.mod is read for its require lines, direct and // indirect. A require states a minimum, so a higher version the build selects before go mod tidy is not seen. A replace to another module and version is judged as that module, one to a local path drops the module, and exclude lines are ignored. go.sum, go.work and vendor/modules.txt are not read, so a module with no go.mod beside it is not judged. A go.mod older than go 1.17, or with no go line, which Go reads as 1.16, lists only part of the build: a finding on one says so, and the modules it leaves out are not judged. A go.mod is read wherever it sits, testdata included. A module path compares as written, case included, and /v2 is a different module from its base, so a record that lower-cases a path, or names a package inside a module, matches nothing. An NVD record that names Go only by collectionURL is not read. A declaration without a version is not judged (an MCP config launch without one is AAK-KNOWN-VULN-003's), and a range is read on release triples, as the pin table reads ranges, so a pre-release bound counts as its release. A range that admits affected and unaffected releases is one INFO finding naming at most five advisories, and none when a lockfile in the same directory resolves the package. A PYSEC record and its GHSA twin are one advisory, reported under the GHSA id; two GHSA records are never merged. Where another rule reports every CVE id of the advisory on the same line of the same file at or above this finding's severity, as most version pins do, only that rule's finding is kept; an advisory without a CVE id is always reported. A lockfile is read for its first 64 distinct releases of a package. NVD's records are read by aak update for the last 120 days of modifications (--nvd-days), and each run's entries replace the last run's, so a record last modified before that window leaves the feed. A record names a package only through its CNA's pkg:pypi, pkg:npm or pkg:cargo packageURL, a PyPI or npm collectionURL (a package's own page, or the registry's root with a packageName, which must agree with a page's name), or a row of the reviewed data/cna_package_map.json matching its repository or its vendor and product (four rows on 2 Nov 2026), never through a name alone; one that maps to none of these is dropped and counted. A CNA version written as text is read in GitHub's advisory grammar (< 0.5.0, >= 2.2.0, < 3.0.1, one lower and one upper bound at most, or = x alone), and > x only on npm, where semver has a first version after x. NVD's CPE data, other text, a version type other than semver, custom or python, and a defaultStatus: affected whose exceptions no interval states are not read; an NVD entry without a CVSS 3.x vector takes its CVSS 4.0 label, else MEDIUM; and one whose CVE an OSV entry for the package names is dropped for OSV's. The bundled records are the two OSV advisories of 27 and 28 Oct 2026 against vLLM and MLflow, the two against anyio below 4.14.2, and nine entries from NVD's records of seven CVEs of 29 Oct to 2 Nov 2026 against langflow, lmdeploy, prefab-ui, open, @modelcontextprotocol/inspector, opencode-ai and @samanhappy/mcphub, so without aak update nothing else is known; a feed over 60 MB keeps only HIGH and CRITICAL records and says so. .egg-info installs, node_modules, conda metadata and image layers are not read. A manifest over 2 MB or a lockfile over 32 MiB is not read.

References


All rules.