Please report security issues privately, not in public issues.

Use GitHub's private vulnerability reporting (Security → Report a vulnerability).

I aim to acknowledge reports within one week.

This policy covers vulnerabilities in pkgaudit itself — for example, a crafted package or tarball that causes pkgaudit to write outside its extraction directory or execute code while scanning.

pkgaudit does not execute the code it analyses. If you believe you have found a vulnerability or malicious code in some other R package, that is not a pkgaudit issue: contact that package's maintainer, and CRAN (cran@r-project.org) if it is distributed there.

The package under audit is untrusted input. pkgaudit's job is to read it without being subverted by it, and to say plainly what it did and did not read.

pkgaudit does not install, build, load, or source the package it analyses. R code is parsed, never evaluated; other files are read as text. A malicious package is therefore malformed input to a reader, not code running in the session doing the scan.

Nothing is fetched, resolved, or reported anywhere. A scan of a package on disk is complete offline, and no information about the package under audit leaves the machine.

audit_tarball() extracts the tarball to a temporary directory (temp_dir, default tempdir()) and removes it when finished. export_unscanned() writes the code pkgaudit cannot read into a directory the caller names; it has no default directory, so naming one is how the caller consents. Directory scans (audit_package()), rule loading (load_rules()) and SARIF output (emit_sarif(), which returns a string) do not write at all.

Validation is fail-closed: an archive that fails any check is refused whole rather than partially extracted. Refused are non-gzip compression (bzip2, xz, zstd, compress — judged by magic bytes, not filename), link entries, GNU long-name and PAX typeflags, path traversal, absolute and drive-qualified paths, paths containing backslashes or control characters, empty paths, decompression bombs (entry-count, size, and expansion-ratio caps), unparseable size fields, and archives without exactly one top-level directory. After extraction the directory is re-checked for symlinks as defense in depth.

Both the content and the file names written by export_unscanned() come from the package under audit, so every target path is resolved and required to lie under the named directory, and a path component that is ., .., or that contains a separator or a control character is refused. Symlinks are never followed, and content is read and rewritten rather than copied, so a link pointing out of the package cannot pull a file in. Nothing is written executable, and nothing is deleted or replaced unless overwrite = TRUE. A refused file is recorded in the returned manifest rather than dropped.

The database is integrity-checked against its SHA-256 sidecar, but the check is time-of-check to time-of-use — see "Security considerations" in ?load_rules. Load rules only from a path whose writers you trust.

audit_tarball() writes the extracted package to temp_dir, and export_unscanned() writes package contents to the directory it is given. If the package under audit may contain sensitive material, or you operate under data-at-rest controls, set temp_dir to an encrypted volume or a tmpfs (RAM-backed) mount. Cleanup uses unlink(), which is not a secure wipe, and a killed process (e.g. SIGKILL) can leave the extraction directory behind.

Findings are for human review. There is no severity model and no pass/fail result: every result in SARIF output carries the same level, and rule coverage is not exhaustive. A package with no findings has not been cleared.

A path assembled at runtime, a name built by paste0(), or a chunk option computed at render time is not followed. pkgaudit also does not build call graphs, so whether a package actually calls a particular function is not something it determines — which is why the phases of code inside a function definition are reported as a stated reading rather than a finding.

pkgaudit reads R and shell. It does not read compiled code or other languages; export_unscanned() hands those to a tool that does. Files are identified by name, not by contents, so an extensionless script — tools/build opening #!/bin/sh — earns no coverage row. The coverage frame reports what each file was read as (parsed, matched as text, exportable to another tool, unexamined, or attempted and failed) so that a clean result can be checked rather than trusted.

The scope of hash_manifest() depends on the exclusion patterns, and symlinks and empty directories are not represented. Prefer auditing the tarball when provenance matters.