The traefik-binary classifier regex required a NUL byte or replacement
character immediately before the version string. On some architectures
(e.g. s390x), the real version string in the binary has no such prefix,
so the classifier instead matched an earlier, unrelated NUL-prefixed
numeric string elsewhere in the binary (e.g. '2.7.0' from an unrelated
printf-style format constant), producing an incorrect version.
This adds an additional pattern, tried first via MatchAny, that matches
version strings terminated by two NUL bytes without requiring a NUL
prefix. The original pattern remains as a fallback for older traefik
binaries (e.g. v1.x) that use single-NUL termination.
Fixes#4980
Signed-off-by: Mayank Sekhar <mayankshekharsingh@gmail.com>
* fix: scan a Java resource adapter (.rar) given as a file source
The file source picks an extractor for the scan target from its file name
alone. A Java resource adapter archive (JCA) uses the .rar extension but is
a zip file, so it was handed to the RAR extractor. With rardecode v2.2.0 the
RAR extractor does not return on non-RAR input (an infinite loop in its
signature search), so `syft scan --from file adapter.rar` never finished.
The same bytes named .jar scan in under a second.
When the name says RAR but the content is a zip archive, leave the file
unextracted, the same way .jar, .war and .ear scan targets are handled. The
java cataloger already reads .rar files (anchore/syft#4136) and now finds the
resource adapter and its nested jars. Real RAR archives are still extracted.
Signed-off-by: Ismael Donmez <ismael.donmez@chainguard.dev>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* chore(deps): bump github.com/nwaples/rardecode/v2 from 2.2.0 to 2.4.1
rardecode v2.2.0 loops forever in its RAR signature search when the input
does not start with a RAR signature (fixed upstream in nwaples/rardecode
f501ac8, "fix infinite loop in findSig", released in v2.2.2). syft reaches
that search for any file source whose name contains ".rar", so a scan of a
self-extracting RAR, or of any other file with such a name, never finished.
With v2.4.1 a self-extracting RAR is extracted, and a file that is not a RAR
archive fails fast with "rardecode: RAR signature not found", the same way a
file named .zip that is not a zip archive fails today.
github.com/mholt/archives v0.1.5 (the latest release) still requires v2.2.0,
so the newer version is selected here.
Signed-off-by: Ismael Donmez <ismael.donmez@chainguard.dev>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix: clarify the .rar zip log line and cover self-extracting RAR sources
- the debug line no longer calls every `.rar`-named zip a resource adapter
- a self-extracting RAR (stub before the signature) given as a file source is
still extracted
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: Ismael Donmez <ismael.donmez@chainguard.dev>
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Alex Goodman <wagoodman@users.noreply.github.com>
SPDX license list 3.29.0 rewrote every opensource.org seeAlso URL from
opensource.org/licenses/<ID> to opensource.org/license/<ID> (and e.g.
opensource.org/license/mit/ to opensource.org/license/MIT). Package metadata
in the wild (Maven POM <license><url>, npm, etc.) still uses the older forms,
so regenerating the list would stop LicenseByURL from resolving them.
Fall back to a normalized index of the opensource.org entries that ignores the
www. prefix, /licenses/ vs /license/, case, and a trailing slash. The index is
built from the generated urlToLicense map, so it works with both the current
and the updated license list without changes to the generator. URL forms that
would normalize to different licenses are left out rather than guessed.
Signed-off-by: Will Murphy <willmurphyscode@users.noreply.github.com>
* ci: drop gen-5 instance families from the test runner
m5a runners run the unit suite ~2.3x slower than m8id (15m vs 6.5m for `go test`), and when spot capacity for newer families dries up runs-on falls back to them. Keeping gen-6 through gen-8 leaves a wide spot pool without the slow tail.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* test: shrink decompression bomb fixtures
The bomb and heap-bound tests pushed 100MB-16GB through gzip and parsers, which under `-race` made them the slowest tests in the unit suite. Each now uses the smallest fixture that still separates the bounded path from the unbounded one.
- arch mtree caps are measured at a small cap and scaled to the shipped constants, so raising a cap still fails the budget
- snap and debian bombs expand to 32MB instead of 512MB (4x their 8MB budgets)
- golang ELF bombs still declare 256MB but deliver 64MB
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* test: add windows unit test job and capture dir resolver path bug
- add a `Unit tests (Windows)` job running a new `make unit:windows` task, scoped to the packages that own host path handling (tests needing linux docker fixtures are skipped)
- `Unit tests` is now a gate job over the Linux and Windows unit jobs
- add a resolver test asserting that locations from `AllLocations` resolve again via `FilesByPath` and use posix paths. This is expected to fail on windows today, which is why file digests come back empty for `dir:` scans (#5325)
- fix windows-only test files that no longer compiled
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* chore: drop verbose output from integration tests
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix: windows path handling outside of the file lookup bug
- chroot paths are always posix (`ToChrootPath` no longer returns backslash paths on windows), same for the directory source name derived from `base`
- `windows.FromPosix` no longer panics on input that isn't volume-encoded (e.g. relative or already native paths)
- skip tests on windows that rely on symlink fixtures, shell script fixtures, or emulate windows paths on a posix host
- disable `core.autocrlf` on the windows CI job so fixture digests match linux
This leaves only the `FilesByPath` lookup failures from #5325.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* test: skip posix-only fileresolver tests on windows
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix: split contained paths on posix separators on windows
- `allContainedPaths` cleaned with the native separator but split on `/`, so a root reached through a symlink never had its ancestors indexed on windows
- index assertions in tests now key on the posix path the index actually stores
- skip symlink fixture tests and a posix-only test helper on windows
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* chore: revert unrelated go.sum churn and allContainedPaths change
- restore `go.sum` and `.make/go.sum` (accidentally populated by a local `-mod=mod` invocation)
- revert the `allContainedPaths` separator change: production passes native windows paths there, and splitting them on `/` produces drive-relative ancestors, so skip the posix-only test on windows instead
- annotate the remaining `filepath` usages that operate on native paths
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix: search the file tree with posix paths on windows (#5325)
The directory and file resolvers key their file tree on posix paths (`/c/some/path`), but `FilesByPath`, `HasPath` and root-anchored `FilesByGlob` searched it with the native path from the chroot context (`C:\some\path`). On windows every lookup silently missed, so anything that resolves files by path found nothing. This is why file digests were empty for `dir:` scans.
Requests are now converted to posix before searching the tree.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(spdx): write a name for the root package when the source has none
A source read from another SBOM (a purl list, a CycloneDX document
without metadata.component, an SPDX document without a DESCRIBES
relationship) can have no name. The SPDX encoder then wrote the document
root package with an empty PackageName, which SPDX requires. In
tag-value output the PackageName line is omitted, so the root package's
fields land in the previous section and syft cannot read its own output
back ("received unknown tag SPDXID in CreationInfo section").
Fall back to the document name for the root package, and make the
document name fall back to "unknown" for every source type rather than
only for unrecognized ones.
Signed-off-by: Tony <tjagodka@gmail.com>
* fix(spdx): drop syft's root package on decode for every SPDX version
- the root package name now comes from one fallback chain (source name, image repo name, source ID, document name), so the root purl no longer gets an empty name for a source decoded from CycloneDX
- decoding strips the root syft writes for an unknown source, so `syft convert` round trips stop adding an `unknown` package each pass
- for SPDX 2.1 and 2.2, which have no `primaryPackagePurpose`, the purpose is taken from the source type in syft's `DocumentRoot-<type>-` ID
**behavior change:** decoding a syft-written SPDX 2.1/2.2 document no longer returns the root as a package or its `CONTAINS` relationships, matching 2.3.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: Tony <tjagodka@gmail.com>
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
Co-authored-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(python): link uv.lock dependents to the locked version they name
A forked uv.lock can lock several versions of one package, and uv records the version on each dependency entry in that case. The parser dropped that field, so every pandas version depended on every numpy version. Carry the locked version through provides and requires as name@version (bare name stays for unique locks) and keep marker text unchanged.
Fixes#5340
Signed-off-by: Dev M <devtechedge@gmail.com>
* fix(python): pair uv.lock versions without widening public metadata
- the locked version on a uv dependency entry is only needed to pair relationships, and the relationships already carry it, so it stays on the internal toml struct instead of `pkg.PythonUvLockDependencyEntry`. The previous approach hid it from JSON but still hashed it into package IDs.
- dependency specs are now built from the raw lock entries at parse time rather than read back off package metadata
- optional-dependencies get the same version pairing, so an extra that names one locked version no longer links to all of them
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: Dev M <devtechedge@gmail.com>
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
Co-authored-by: Alex Goodman <wagoodman@users.noreply.github.com>
Motivation:
syft's PHP interpreter cataloger reports each bundled PHP C extension
(openssl, ldap, sqlite3, etc.) as its own package, using the PHP
interpreter's own version number as the package version (since these
extensions are compiled into a specific PHP release and have no
independent version of their own). The classifier already attaches an
explicit CPE like cpe:2.3:a:php-openssl:php-openssl:8.5.7:... to make
clear this refers to the bundled PHP extension, not upstream OpenSSL.
However, that CPE was marked with Source: cpe.GeneratedSource. Syft's
package-finalization pipeline (internal/task/package_task_factory.go)
only skips its generic, name-based CPE generation when a package
already has an "authoritative" CPE (any Source other than
GeneratedSource) - so it also generated and appended a second CPE,
cpe:2.3:a:openssl:openssl:8.5.7:..., derived from the bare package
name. That second CPE is indistinguishable from a real standalone
OpenSSL/LDAP/SQLite installation and can cause vulnerability scanners
that consume this SBOM to falsely flag PHP hosts as running vulnerable
upstream libraries at version "8.5.7" (a PHP version, not a real
OpenSSL/LDAP/SQLite release). See anchore/syft#5014 for a full repro.
Approach:
Mark the classifier's explicit CPE with Source: cpe.DeclaredSource
instead of cpe.GeneratedSource, so it is treated as authoritative and
the extra name-derived CPE is no longer generated. This mirrors the
existing pattern in
syft/pkg/cataloger/java/parse_jvm_release.go's newJvmCpe, which uses
the same DeclaredSource + comment convention to mark a
syft-synthesized (not literally file-declared) CPE as trustworthy and
suppress further generation. It also brings this classifier in line
with its sibling php-cli/php-fpm/php-apache classifiers in the same
file, which already mark their CPEs as authoritative (via
NVDDictionaryLookupSource, not applicable here since php-<ext> is not
a real NVD dictionary entry).
This change only suppresses the extra generated CPE; it does not
change the package name, type, version, or PURL, and the existing
php-<ext>:php-<ext> CPE is unaffected.
Validation:
go build ./...
go test ./syft/pkg/cataloger/php/... ./internal/task/... ./syft/pkg/cataloger/internal/cpegenerate/...
All pass except the two pre-existing docker-image-fixture subtests of
Test_InterpreterCataloger, which fail in this sandbox only because the
`docker` binary isn't available (exec: "docker": executable file not
found in $PATH); confirmed via `git stash` that these two subtests
fail identically on main without this change, i.e. this is a sandbox
limitation and not a regression. Added a new unit test,
Test_getClassifier_declaredCPESource, that directly asserts the
classifier's CPE now carries Source: cpe.DeclaredSource.
Fixes#5014
Signed-off-by: Pujitha Paladugu <10557236+pujitha24@users.noreply.github.com>
Co-authored-by: Pujitha Paladugu <10557236+pujitha24@users.noreply.github.com>
* fix: report scanner errors instead of returning a partial parse as complete
A line longer than `bufio.Scanner`'s 64KB buffer ends the scan with `ErrTooLong`, which only `Err()` reports. Parsers now return what they found along with the error (recorded as an unknown for the file), or log it where there is no error to return.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix: allow lines up to 1MB in line-oriented package metadata parsers
Adds `internal.NewLineScanner`, a `bufio.Scanner` capped at 1MB per line rather than the 64KB default, the same cap the alpm parser already uses. Long single-line fields (descriptions, license text, dependency lists) now parse instead of ending the scan early.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix: use io.ReadFull where a short read was silently accepted
- dotnet bundle 7-bit int decoding could read a zero-byte result as a zero byte
- wordpress plugin header parsing included the unread tail of its buffer
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix: finish moving line scanners to the 1MB cap and checking scanner errors
- java manifest, pom.properties, Gemfile.lock, requirements.txt, homebrew formula, alpm and the purl/cpe decoders now use `internal.NewLineScanner`
- snap kernel changelog reports a scan error instead of silently dropping the base kernel package
- the attestation output reader keeps draining the pipe after a scan error so cosign can't block on a full pipe
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix: stop logging the full setup.py line once per dependency match
The debug log ran once per regex match and included the whole line, so log volume grew with the square of line length. With 1MB lines that is hundreds of GB of debug output from a single crafted file.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(gentoo): don't dereference a nil location when a LICENSE scan fails
`extractLicenses` accepts a nil location, but the scan error log added earlier on this branch dereferenced it.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(gentoo): bound license group expansion on hostile license_groups
- dedup within a group uses a set instead of `slices.Contains`, which was quadratic in tokens per line and got much worse under the 1MB line cap
- each group is expanded once and memoized, so a chain of groups that reference each other repeatedly no longer expands exponentially
- repeated references to the same group within one line are skipped
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(javascript): don't panic on a yarn v1 lockfile line missing its value
A v1 `yarn.lock` dependency line without a version constraint, or a
`version`/`resolved`/`integrity` line without a value, indexed past the end of
the split line. Such dependencies are now skipped and empty fields left empty.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(alpine): don't panic on an apk db ACL or checksum line before its entry
An `M:`, `a:` or `Z:` line that comes before any directory or file entry in
`lib/apk/db/installed` indexed `files[-1]`. Such a line is now skipped.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(erlang): don't panic on a truncated rebar.lock or .app file
The term parser read past the end of the input when a file ended mid-term:
whitespace only, an unclosed list, a lone `<`, or a comment ending in `\r`.
These now return an "unexpected end of input" error.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(lua): don't panic on a truncated .rockspec
The rockspec parser read past the end of the input when a file ended mid-statement:
a lone `[`, an unclosed `[[...]]` value or `[...]` index, or a `local` assignment cut off
before its `=`, value, or line end. These now return an "unexpected end of input" error.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(swift): don't panic on a Podfile.lock pod without a version
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(gentoo): don't panic on a short obj line in portage CONTENTS
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(spdx): don't panic on an opam checksum without an algorithm
The opam cataloger stores checksums verbatim, so a malformed one reached the SPDX encoder and panicked there, outside the cataloger's recover. Checksums without an `=` are now skipped.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* refactor: use strings.Cut where a split was only indexed after a Contains check
Also guards `windows.FromPosix` against a path with no separator.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* test: guard untyped package metadata fields
- new `any`-typed metadata fields must be allowlisted with a reason
- allowlisted untyped fields are pushed through `SetID` and every encoder with
non-finite, deeply nested and large values
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* test: fuzz every generic cataloger parser
One fuzz target over every glob-selected parser in the capabilities data, run
through its cataloger task. Each input must not panic, must finish in time and
must keep allocations proportional to its size. Seeds come from each parser's
testdata plus the string literals in its package.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* test: scan malformed inputs under memory and time limits
A directory of small malformed files (deep nesting, oversized declared counts,
truncation, type mismatches, non-finite floats) is scanned by the CLI with a
memory limit and an RSS ceiling. The scan must exit 0 and write an SBOM.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* test: stop a fuzz input once its live heap passes the allocation limit
The allocation check only ran after the parser returned, so an input on its way to exhausting memory ran the machine into swap first. The harness now polls the live heap while the parser runs and panics past the same limit, which the fuzz engine saves as a crasher.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix: bound nesting depth and error context in the erlang and lua parsers
- a deeply nested rebar.lock or .rockspec overflowed the goroutine stack, which is fatal rather than a recoverable panic, so one file could take down the whole scan. Both parsers now stop at 1000 levels with an error.
- parse errors echoed whole input lines into the SBOM as unknowns, and lua appended another copy at every nesting level. `PrintError` now clips to a window around the error column, and lua decorates once at the top.
- the Podfile.lock error no longer echoes the pod entry.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* test: close ways the malformed-input checks could pass quietly
- a ctx-aware parser hitting the fuzz timeout now fails the input instead of racing the hard deadline
- the heap watchdog measures growth from a post-GC baseline rather than an absolute process-wide number
- a saved fuzz input naming a target that no longer exists fails instead of being skipped
- the CLI malformed-input scan fails on a recovered parser panic, not only on a non-zero exit
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(cpe): cap sub-selection prefixes to bound candidate cross-product
a name with thousands of `-`/`_` separators produced one vendor and product candidate per segment, and the vendor x product cross-product made CPE generation roughly cubic in time and memory (8KB name: ~250s, 4.8GB RSS). `generateSubSelections` now stops after 10 prefixes. The full name is still a candidate, so real package names are unaffected.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(cpe): bound candidate count and field length before the cross-product
the prefix cap only covered one way to inflate the vendor x product cross-product in `FromPackageAttributes`. Java groupIDs add a vendor (and, with no artifactID, a product) per dot segment, and .NET executables add one of each per file, so a 1000-segment `Bundle-SymbolicName` still produced ~1M CPEs and several GB.
- vendor and product candidates are each capped at 64 (shortest first, so truncation is deterministic)
- names, versions, and candidates over 512 bytes are skipped, since every byte is paid once per CPE in key building and regex validation
real packages produce around a dozen candidates with names well under 100 bytes, so output for real inputs is unchanged. Also corrects the earlier comment: prefixes only become vendor candidates and the cost was quadratic, not cubic.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
UPX pads to a 4 byte boundary before it writes the loader stub. The
image-small-upx build happened to need no padding, so the cataloger
tests over it passed with or without the fix in #5347.
- build the fixture with -X main.Version=1.0.11, which needs 2 bytes of
padding, so the existing cataloger tests fail without the fix
- add TestImageSmallUPXNeedsLoaderPadding, which fails if a change to the
fixture's inputs means it no longer needs padding
- run the crafted loader test on both sides of the boundary, so a skip
that always rounds up is caught too
Signed-off-by: Will Murphy <willmurphyscode@users.noreply.github.com>
UPX pads its output to a 4 byte boundary before it writes the loader
stub, and l_lsize counts from that boundary. The block chain resumed at
the end of the last PT_LOAD extent plus l_lsize, so whenever the
compressed extents ended off a 4 byte boundary it landed 1 to 3 bytes
early, read a misaligned b_info, and stopped. The padding between the
segments was never filled in, the reconstruction ended at that first
hole, and .go.buildinfo in the data segment went with it, so the binary
contributed no packages.
The loader skip now rounds up to 4 before adding l_lsize, which matches
the funpad4 in UPX's own unpacker. The crafted fixture for the tail
extents put its stub straight after the head extent, the layout the bug
assumed, so it now pads the way UPX does.
Signed-off-by: huuyafwww <huuya.yamauchi@3-shake.com>
Same issue as #5350: go1.27's compress/flate encodes this 512MB fixture to just over 1MB, so the fixture sanity check tripped before the bound was exercised. The line bound itself is fine.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
go1.27's compress/flate encodes 512MB of zeros to ~1MB (1052690 bytes) where go1.26 gives ~510KB, so the fixture sanity check tripped on newer toolchains before the bound was ever exercised. The decompression bound itself is fine (bounded read allocates ~300KB on both).
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
- pre-scan the GGUF metadata KV section and reject values gguf-parser-go would panic on, recurse on without bound, or overallocate for: wrong typed well-known keys, zero alignment, nested or oversized fully decoded arrays, an unreasonable `split.tensors.count`, and empty strings the parser would misread. Keys are matched the way the parser sees them. The KV count, fully read string bytes, and skipped string counts are bounded too. The file becomes an unknown instead of taking down the scan.
- `copyHeader` now bounds its own read to the max header size.
- non-finite floats in `header` (including inside arrays) become `"NaN"`, `"+Inf"` or `"-Inf"` so JSON encoding can't fail. `header` and `metadataHash` are otherwise unchanged.
- a `metadataHash` failure is now an error instead of an empty hash.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(ai): keep merging GGUF layers when one fails to parse
a single unparseable layer in an OCI GGUF artifact skipped the merge, so the
named model lost its parts and nameless layers were dropped. The error is
still returned.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(ai): only send GGUF layers to the GGUF cataloger
the media type glob `application/vnd.docker.ai*` also matched config,
model.file, license and safetensors layers, so every safetensors OCI model
reported spurious "invalid GGUF magic number" unknowns. Match
`application/vnd.docker.ai.gguf.v3` exactly.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(ai): don't wrap safetensors parameter counts
- tensors whose shape product overflows uint64 are skipped instead of wrapping
- per-file and per-shard parameter totals saturate instead of wrapping
- parameter count and dominant dtype are computed in one pass
- GGUF licenses are created with the scan context, and drop no-op unknown checks
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* refactor(ai): derive safetensors model info from the header in one place
the file and OCI layer parsers built identical metadata from a parsed header.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(ai): keep the GGUF media type with the GGUF cataloger
also corrects stale comments in the GGUF merge processor.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
Companion files next to AI models (`config.json`, `README.md`, and the OCI model.file, license and model config layers) are now read with named size limits, and are only used when they look like what they claim to be.
- JSON companions are read whole up to a limit and skipped when larger, rather than being fully buffered or parsed truncated. OCI companion layers are fetched one byte past the cap so oversize is detectable.
- README and license layers read a prefix. The frontmatter block is capped at 256 KB and list values at 32.
- dir and OCI scans share one 8 MB safetensors header cap. Directory scans used to allow 100 MB.
- `config.json` is only used when it has `architectures` or `_name_or_path`, and is looked up beside the model or in its parent directory, no longer all the way up to the scan root
- OCI model.file layers are processed in a stable order, so the chosen name no longer changes between runs
- README frontmatter:
- a list-form `license:` is accepted and no longer drops `base_model`
- `null` values are ignored
- delimiters must be exact `---` lines
- model names from `_name_or_path` / `base_model`:
- Windows paths are handled
- `.`, `..` and `/` fall back to the next name source
- names are capped at 256 characters
Behavior changes worth noting:
- a `config.json` two or more directories above the model no longer names it
- JSON `config.json` files without `architectures` or `_name_or_path` (e.g. `generation_config.json`, app configs) are ignored
- safetensors headers over 8 MB in directory scans are now rejected, matching OCI scans (real shard headers are well under 1 MB)
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(generic): recover from parser and processor panics
- a panic while parsing one file becomes an unknown for that location instead of failing the whole SBOM
- a panic in a package processor is logged as a warning and the unprocessed results are kept
- stacks go to the trace log, not into the unknown text
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(unknown): keep errors that unwrap to nil
`sync.PanicError` with a non-error panic value (e.g. `panic("boom")`) unwraps to nil, and was being dropped entirely from the error graph instead of being kept as a leaf.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(task): don't put the stack trace in task panic errors
Task panics are still fatal, but the error now reads `panic in task "<name>": <value>` and the stack goes to the debug log.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* test(generic): cover chained processors when a later one panics
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(file): recover from panics in the executable and file digest catalogers
A panic while reading one file is now recorded as an unknown for that file instead of failing the whole SBOM, the same as generic catalogers.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(ai): use go.yaml.in/yaml/v3 to parse model card frontmatter
gopkg.in/yaml.v3 panics on some malformed merge-key payloads (for
example "<<:\r? 0:"), which crashed the gguf and safetensors catalogers
when a model README contained such frontmatter. Switch the Hugging Face
frontmatter parser to the maintained go.yaml.in/yaml/v3 fork, which is
already used by the other catalogers and returns an error instead of
panicking. The existing error path logs and skips the frontmatter.
Fixes#5327
Signed-off-by: tk1475 <118763009+tk1475@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* chore: finish moving off gopkg.in/yaml.v3
- move the snap catalogers and `internal/capabilities` to `go.yaml.in/yaml/v3`. The snap parsers decode yaml from scanned sources and hit the same merge-key panic as #5327
- enable `depguard` with a deny rule for `gopkg.in/yaml.v3` so it can't be reintroduced
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: tk1475 <118763009+tk1475@users.noreply.github.com>
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Alex Goodman <wagoodman@users.noreply.github.com>
The package supplier is optional in SPDX, but fileSource dereferenced it
without the nil check containerSource already has, so a valid SPDX 2.3
document whose described package has primaryPackagePurpose FILE and no
supplier crashed syft convert and grype sbom: input.
A JSON null in files, a package's files or externalRefs (or an SPDX 2.1
package the version converter cannot map) decodes to a nil pointer that
the conversion then dereferences. Drop nil entries once at the start of
ToSyftModel instead of guarding each loop.
Signed-off-by: Tony <tjagodka@gmail.com>
- log lines shown under the TUI were glued onto the last UI row, and the last row could get erased (and overwritten by the report) when the TUI exits. Fixed in `bubbly` v0.2.2
- log colors were lost because the logger latched its TTY check on the first entry, which went to the TUI footer buffer. `go-logger` v0.2.0 checks per entry and supports `TerminalWriter`, which the footer writer now implements (`NO_COLOR` is still honored)
- log lines under the TUI get a dim `Logs (non-fatal, see -v for more):` title with a branch pointing at the first entry, so it's clear what the trailing lines are
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
Docker Hub rebuilt and re-pushed `busybox:1.38.0`, which changed the per-platform config digests and broke `TestMultiPlatformOCIImageSelection` on uncached runs.
- pin the fixture to the busybox 1.38.0 index digest so future re-pushes can't shift it
- update the expected per-platform digests to match
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* fix(java): resolve Groovy 4+ group IDs to org.apache.groovy
Groovy 4.0 moved from the Codehaus coordinates (org.codehaus.groovy) to the
Apache Software Foundation (org.apache.groovy). JARs that ship no pom
metadata fall through to the static artifact-to-group map, which carries the
3.x-era org.codehaus.groovy entries for the whole groovy family, so 4.x
artifacts were attributed to an obsolete group even though their manifest
identifies the Apache vendor.
Where the static map attributes an artifact to the historical group, a
major-version migration resolves groovy 4.x and newer to org.apache.groovy,
while 3.x and earlier keep their historical group. The Eclipse-hosted
groovy-eclipse-* artifacts, which share the prefix but kept their Codehaus
coordinates, are excluded.
Fixes#5311
Signed-off-by: texasich <101962694+texasich@users.noreply.github.com>
* refactor(java): share version-aware groovy group ID lookup with CPE generation
- move the artifact-to-group lookup behind `cpegenerate.ArtifactIDToGroupID(artifactID, version)` so the PURL and CPE paths agree; the exported mutable map is now unexported
- version-dependent entries are expressed as a `groupIDMigration{before, after, atMajor}` (groovy moved to `org.apache.groovy` at 4.0); unconditional entries, including `groovy-eclipse-*` and apache-only 4.0+ modules, stay in the plain map
- add apache-only modules (`groovy-ginq`, `groovy-contracts`, `groovy-callsite`, ...) that otherwise took the group from their per-module `Automatic-Module-Name`
This changes CPEs for pom-less groovy 4+ jars (all real ones): vendors are now `apache` (e.g. `cpe:2.3:a:apache:groovy:4.0.33`) instead of `codehaus`, matching the PURL, jars with pom metadata, and NVD.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: texasich <101962694+texasich@users.noreply.github.com>
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
Co-authored-by: Alex Goodman <wagoodman@users.noreply.github.com>
closes#5299: docker:// references are OCI images, not GitHub repos.
Returning empty string produces a package with no PURL, matching the
treatment for local composite actions.
Signed-off-by: Yunare Maia <yunare@gmail.com>
An extra-dep can be pinned as "@sha256:<hash>,<size>" or as "@rev:<n>". The hash
slice assumed the first form, so a rev pin left the end index at -1 and the
parser panicked with slice bounds out of range. A rev pin carries no package
hash, so only the sha256 form now yields one.
Signed-off-by: Arpit Jain <arpitjain099@gmail.com>
* build(deps): bump go-make to v0.8.1
Picks up golangci-lint 2.13.2 (built with go1.27) in place of 2.11.4 (go1.26).
The old build panics outright on a go1.27 GOROOT with "file requires newer Go
version go1.27", and an older mise-shim build fails the quieter way: it loses
type info and reports zero findings with exit 0, so a ruleguard run comes back
green without having evaluated anything.
Also folds in what the newer linter flags:
- `reflect.Ptr` -> `reflect.Pointer` in three files, applied by --fix
- `tw.CellFormatting{MergeMode: ...}` -> `tw.CellMerging{Mode: ...}` in the
cataloger info table, for a staticcheck SA1019 deprecation
Note the deprecation text names `CellConfig.CellMerging.Mode`, but the field is
`CellConfig.Merging`; `CellMerging` is the type. That command path has no test
coverage, so the swap was checked by rendering the same hierarchical-merge table
both ways and diffing: byte identical.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* chore(lint): tune goconst for data tables
goconst got stricter in golangci-lint 2.13 and now counts strings in map
literals and struct fields, which lights up every lookup table in the repo. It
was already enabled, so this is main going red on a linter bump rather than
anything new in the tree: 101 findings, and the default threshold of 3 means a
vendor name appearing in three rows of a table reads as a magic constant.
Raising `min-occurrences` to 5 takes that to 14. The rest split by whether the
repeated string is data or an actual constant:
- `cpegenerate` is excluded by path. It maps artifact names to vendors,
products, and maven group ids, and repeats strings 35, 83, even 293 times, so
no threshold reaches it. Hoisting those would trade a readable table for
indirection.
- the two SBOM fixture builders in `format/internal/testutil` get a
function-level nolint each. The repeated names and versions are what the
golden files assert against.
- `mysql-cluster` and `php` were genuine repeated literals and are now
constants.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* chore(test): extract measureAlloc into a shared helper
The allocation-budget tests added during the unbounded allocation work grew a
copy of the same helper in every package they touched: four identical
`measureAlloc` definitions plus two more spelled out inline against
`runtime.MemStats`.
Now `testutils.MeasureAlloc`, at the repo root so `internal/spillbuf` can reach
it too. The doc comment carries the two things that were only recorded in some
of the copies: TotalAlloc is process-wide so callers must not use t.Parallel,
and a budget assertion only means something if the fixture would really blow
past it unguarded, so measure the unguarded path as well.
No behavior change, the body was identical in all six.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* chore(lint): add ruleguard rules for overflow and swallowed EOF
Two rules out of the recent unbounded allocation work (#5187, #5195, #5267,
#5291, #5292, #5293, #5294).
`noOverflowingBoundsCheck` flags `offset + size > limit` (and <, <=, >=) where
an operand is unsigned, neither addend is constant, and the operands are at
least 32 bits. Unsigned addition wraps, so an offset out of a file header can
carry the sum past the wrap point and pass a check it should have failed. The
fix is to subtract against the limit instead. Constant addends are excluded
because `pos+4 <= uint32(len(buf))` is the normal way to step a cursor, and the
32-bit floor drops nibbles widened for a comparison. Would have caught four
sites across #5195 and #5291.
`noSwallowedEOF` flags `return nil` / `return nil, nil` on an io.EOF check. The
caller gets a zero value it cannot tell from real data, and anything decoding a
fixed-size struct off it parses the padding as content. That is the shape behind
#5186.
Both decide from the expression alone, and writing the fix deletes the pattern,
so neither needs a suppression anywhere in the tree. That was the bar: rules for
the allocation and decompression classes were tried and dropped, because the
bound they look for lives in a preceding statement (or, for a decompressor
constructor, always downstream), so no correct implementation can silence them.
That would have meant ~15 permanent nolints on code that is already right. The
rules here that do work don't judge whether a bound exists, they route to a
wrapper that takes one: noDirectTempFiles, noDirectELFOpen.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* test: pin allocation budgets on the alpm, deb, and changelog bounds
The bounds added for #5292, #5293 and #5294 are asserted by the error they
return. That keeps passing if the bound is removed and something else rejects
the input later, which is the failure these were written for. Each of these
measures bytes instead, and proves the fixture is really a bomb first so a
fixture that quietly stops being one fails rather than passing.
- alpm mtree: 16MB of newlines, well inside the 64MB byte cap, costs 16.5GB
uncapped against 238MB with maxMtreeLines. That 16.5GB is the measurement the
line cap was added for and it was recorded only in the commit message. Note
what the cap actually buys: 300k entries still cost a few hundred MB, per
concurrent cataloger.
- deb members: 1.28GB against 392KB. Both arms read with io.ReadAll on purpose,
since the stream itself never holds much whatever the cap says; what the cap
protects is a consumer that buffers, and the real one is a tar reader building
entries.
- snap changelog: 1.28GB against 175KB. This one has no explicit cap at all. It
is held by bufio.Scanner refusing a token past MaxScanTokenSize, which is
incidental, so switching to io.ReadAll or a bufio.Reader to pick up long lines
would remove it silently.
Nothing added for the kernel module reader: its copy is already
io.Copy(tempFile, io.LimitReader(rc, max+1)), so peak disk is bounded by
construction and heap is streaming. A heap budget would assert the wrong
resource, and the three existing subtests cover the cap, at-cap fidelity, and
spill cleanup.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* test: pin allocation budgets on the ai and alpm byte caps
Follows the same pattern as the alpm line cap: measure bytes, and prove the
fixture is really a bomb first so one that quietly stops being a bomb fails
rather than passes.
- safetensors: 252MB against 2.6KB. The existing over-cap case deliberately
omits the body, since the guard rejects on the declared length before reading
anything. That is the right test for the guard but it would keep passing if
the check moved after the read, so this one delivers the bytes.
- gguf: 268MB against 168MB. copyHeader io.Copy's whatever it is handed, so
maxHeaderSize holds only because every caller wraps the reader first. This
pins that the limit is where the bound lives, so dropping the wrapper at a
call site fails here.
- alpm byte cap: 740MB against 132MB, the counterpart to the line cap budget.
Worth knowing for anyone writing the next one of these: a single enormous line
does not exercise the alpm byte cap. go-mtree runs its own bufio.Scanner and
refuses a token past 64KB, so that fixture fails as "token too long" before the
byte cap is consulted. The fixture here uses ordinary 1KB lines, enough of them
to pass 64MB while staying under the 300k line cap, so the byte cap is actually
the thing under test.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* test: assert bytes retained instead of allocated for gguf copyHeader
the bounded arm writes into a bytes.Buffer, whose doubling growth costs a
multiple of what it holds and lands on a different power of two depending on
the platform. it measured 168MB locally and 320MB on CI, past the budget.
buf.Len() states the same property deterministically.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* Update steroscope to include OCI platform metadata
Signed-off-by: Jason Paulos <jasonpaulos@users.noreply.github.com>
* chore(deps): point stereoscope at anchore/main
drops the fork `replace` directive and moves to a pseudo-version off stereoscope
main (`a22e5ae`, anchore/stereoscope#657). no tag carries it yet, so this should
move to a real version once one is cut.
`readImageMetadata` now fills os/architecture/variant for every provider, not
just oci. the docker tarball provider set neither before, so image metadata and
the image purl change for docker-archive input too:
- `"architecture": ""` -> `"amd64"`
- `"os": ""` -> `"linux"`
- `?arch=` -> `?arch=amd64`
the empty `arch=` qualifier was never right, but it is user-visible output.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
* test: cover single-platform oci layouts
the pr only added assertions to the multiplatform fixture, which already
reported the right platform. the behavior that actually changed is single-image
layouts: they used to report empty os/arch and silently ignore `--platform`, and
now report what the image config says and fail with a platform mismatch.
also drops the non-linux skip on the default-platform test. the oci providers
default to `linux/<GOARCH>` regardless of host os, so it holds on macos too.
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
---------
Signed-off-by: Jason Paulos <jasonpaulos@users.noreply.github.com>
Signed-off-by: Alex Goodman <wagoodman@users.noreply.github.com>
Co-authored-by: Alex Goodman <wagoodman@users.noreply.github.com>
When javascript.search-remote-licenses is enabled, the lock catalogers
query registry.npmjs.org once per package occurrence, on every run, with
nothing in front of the network call. The same name@version is fetched
again for every lockfile that mentions it, and again on the next scan.
This adds a javascriptLicenseResolver that resolves licenses through
syft's on-disk cache, keyed on "<name>/<version>" and namespaced as
javascript/v1, and routes the package-lock v1, package-lock v2, pnpm,
yarn and bun package constructors through it, leaving
getLicenseFromNpmRegistry itself unchanged. Failed lookups are not
cached: a transient registry failure or a name that 404s must not be able
to blank out licenses for the lifetime of the cache entry, since a
missing license silently weakens downstream license policy checks. A
successful lookup that returns an empty license is a real answer and is
still cached.
Signed-off-by: Anže Pečar <anze@pecar.me>
Mayank Sekhar
anchore-oss-update-botandanchore-oss-update-bot
Will Murphy
Darragh McGurk
KR RavindraandWill Murphy
Antoni (Tony) JagodkaandAlex Goodman
Dev MandAlex Goodman
Pujitha PaladuguandPujitha Paladugu
huuya
Sverrir Sigurdsson
dependabot[bot]
texasichandAlex Goodman
Yunare Maia
Arpit Jain
somaz
Jason PaulosandAlex Goodman
Anže Pečar