This is just preliminary support for identifying an assortment of
different AI model files.
So far, this detects the following types:
- GGML GGUF (.gguf)
- ONNX AI (.onnx)
- TensorFlow Lite (.tflite)
Additional types to consider:
- SafeTensors (.safetensors)
- TensorFlow (.pb, .ckpt, .tfrecords)
- Keras (.keras)
- pickle (.pkl)
- numpy (.npy, .npz)
- coreml (.coreml)
- PyTorch (.pt, .pth, .bin, .mar, .pte, .pt2, .ptl)
Outside of being able to differentiate by file type, the scanner
will treat CL_TYPE_AI_MODEL the same as CL_TYPE_BINARY_DATA.
We're not adding parsers to further process these files, for now.
It may be necessary to differentiate between *.pyc and other binary
types in case additional processing is needed.
Outside of being able to differentiate the by file type, the scanner
will treat CL_TYPE_PYTHON_COMPILED the same as CL_TYPE_BINARY_DATA.
That is - we're not adding parser at this time to further break down
.pyc files.
Includes rudimentary support for getting slices from FMap's and for
interacting with libclamav's context structure.
For now will use a Cisco-Talos org fork of the onenote_parser
until the feature to read open a onenote section from a slice (instead
of from a filepath) is added to the upstream.
Some small files that appear to include archive content (zip entries,
RAR archives, etc) later in the file are no longer set to that SFX type
initially. However, the current way we skip those drops the initial
type detection on the floor, meaning that something like a small image
may end up as CL_TYPE_ANY instead of CL_TYPE_PNG.
This commit fixes that by changing where in the process the SFX types
are ignored.
The `cli_memcpy` function is a version of memcpy that will catch
exceptions on Windows if the memcpy fails because the memory-mapped file
fails to read (like if a network drive is disconnected or something).
It presently returns an `int`. Since it is only used in one spot, I've
swapped it over to a `cl_error_t`, and created a unique variable so
we're not piggybacking with the `cli_file_t` enum used to check the
result of a filetyping scan.
If initial file type recognition comes back as an SFX type, which may
happen for small files that do not get recognized as any other file type
and contain a zip entry somewhere in the middle, then the type will be
set to that SFX type. This is a problem because later on when we go to
do embedded file type recognition, we explicitly skip SFX types, in
addition to TAR's and other types that are parsed elsewhere and have a
high embedded file type recognition FP-rate because they aren't
compressed.
This commit prohibits that initial FTM check from selecting an SFX type.
The SFX type will be rediscovered in `scanraw()` where the type is
handled/parsed.
Scan recursion is the process of identifying files embedded in other
files and then scanning them, recursively.
Internally this process is more complex than it may sound because a file
may have multiple layers of types before finding a new "file".
At present we treat the recursion count in the scanning context as an
index into both our fmap list AND our container list. These two lists
are conceptually a part of the same thing and should be unified.
But what's concerning is that the "recursion level" isn't actually
incremented or decremented at the same time that we add a layer to the
fmap or container lists but instead is more touchy-feely, increasing
when we find a new "file".
To account for this shadiness, the size of the fmap and container lists
has always been a little longer than our "max scan recursion" limit so
we don't accidentally overflow the fmap or container arrays (!).
I've implemented a single recursion-stack as an array, similar to before,
which includes a pointer to each fmap at each layer, along with the size
and type. Push and pop functions add and remove layers whenever a new
fmap is added. A boolean argument when pushing indicates if the new layer
represents a new buffer or new file (descriptor). A new buffer will reset
the "nested fmap level" (described below).
This commit also provides a solution for an issue where we detect
embedded files more than once during scan recursion.
For illustration, imagine a tarball named foo.tar.gz with this structure:
| description | type | rec level | nested fmap level |
| ------------------------- | ----- | --------- | ----------------- |
| foo.tar.gz | GZ | 0 | 0 |
| └── foo.tar | TAR | 1 | 0 |
| ├── bar.zip | ZIP | 2 | 1 |
| │ └── hola.txt | ASCII | 3 | 0 |
| └── baz.exe | PE | 2 | 1 |
But suppose baz.exe embeds a ZIP archive and a 7Z archive, like this:
| description | type | rec level | nested fmap level |
| ------------------------- | ----- | --------- | ----------------- |
| baz.exe | PE | 0 | 0 |
| ├── sfx.zip | ZIP | 1 | 1 |
| │ └── hello.txt | ASCII | 2 | 0 |
| └── sfx.7z | 7Z | 1 | 1 |
| └── world.txt | ASCII | 2 | 0 |
(A) If we scan for embedded files at any layer, we may detect:
| description | type | rec level | nested fmap level |
| ------------------------- | ----- | --------- | ----------------- |
| foo.tar.gz | GZ | 0 | 0 |
| ├── foo.tar | TAR | 1 | 0 |
| │ ├── bar.zip | ZIP | 2 | 1 |
| │ │ └── hola.txt | ASCII | 3 | 0 |
| │ ├── baz.exe | PE | 2 | 1 |
| │ │ ├── sfx.zip | ZIP | 3 | 1 |
| │ │ │ └── hello.txt | ASCII | 4 | 0 |
| │ │ └── sfx.7z | 7Z | 3 | 1 |
| │ │ └── world.txt | ASCII | 4 | 0 |
| │ ├── sfx.zip | ZIP | 2 | 1 |
| │ │ └── hello.txt | ASCII | 3 | 0 |
| │ └── sfx.7z | 7Z | 2 | 1 |
| │ └── world.txt | ASCII | 3 | 0 |
| ├── sfx.zip | ZIP | 1 | 1 |
| └── sfx.7z | 7Z | 1 | 1 |
(A) is bad because it scans content more than once.
Note that for the GZ layer, it may detect the ZIP and 7Z if the
signature hits on the compressed data, which it might, though
extracting the ZIP and 7Z will likely fail.
The reason the above doesn't happen now is that we restrict embedded
type scans for a bunch of archive formats to include GZ and TAR.
(B) If we scan for embedded files at the foo.tar layer, we may detect:
| description | type | rec level | nested fmap level |
| ------------------------- | ----- | --------- | ----------------- |
| foo.tar.gz | GZ | 0 | 0 |
| └── foo.tar | TAR | 1 | 0 |
| ├── bar.zip | ZIP | 2 | 1 |
| │ └── hola.txt | ASCII | 3 | 0 |
| ├── baz.exe | PE | 2 | 1 |
| ├── sfx.zip | ZIP | 2 | 1 |
| │ └── hello.txt | ASCII | 3 | 0 |
| └── sfx.7z | 7Z | 2 | 1 |
| └── world.txt | ASCII | 3 | 0 |
(B) is almost right. But we can achieve it easily enough only scanning for
embedded content in the current fmap when the "nested fmap level" is 0.
The upside is that it should safely detect all embedded content, even if
it may think the sfz.zip and sfx.7z are in foo.tar instead of in baz.exe.
The biggest risk I can think of affects ZIPs. SFXZIP detection
is identical to ZIP detection, which is why we don't allow SFXZIP to be
detected if insize of a ZIP. If we only allow embedded type scanning at
fmap-layer 0 in each buffer, this will fail to detect the embedded ZIP
if the bar.exe was not compressed in foo.zip and if non-compressed files
extracted from ZIPs aren't extracted as new buffers:
| description | type | rec level | nested fmap level |
| ------------------------- | ----- | --------- | ----------------- |
| foo.zip | ZIP | 0 | 0 |
| └── bar.exe | PE | 1 | 1 |
| └── sfx.zip | ZIP | 2 | 2 |
Provided that we ensure all files extracted from zips are scanned in
new buffers, option (B) should be safe.
(C) If we scan for embedded files at the baz.exe layer, we may detect:
| description | type | rec level | nested fmap level |
| ------------------------- | ----- | --------- | ----------------- |
| foo.tar.gz | GZ | 0 | 0 |
| └── foo.tar | TAR | 1 | 0 |
| ├── bar.zip | ZIP | 2 | 1 |
| │ └── hola.txt | ASCII | 3 | 0 |
| └── baz.exe | PE | 2 | 1 |
| ├── sfx.zip | ZIP | 3 | 1 |
| │ └── hello.txt | ASCII | 4 | 0 |
| └── sfx.7z | 7Z | 3 | 1 |
| └── world.txt | ASCII | 4 | 0 |
(C) is right. But it's harder to achieve. For this example we can get it by
restricting 7ZSFX and ZIPSFX detection only when scanning an executable.
But that may mean losing detection of archives embedded elsewhere.
And we'd have to identify allowable container types for each possible
embedded type, which would be very difficult.
So this commit aims to solve the issue the (B)-way.
Note that in all situations, we still have to scan with file typing
enabled to determine if we need to reassign the current file type, such
as re-identifying a Bzip2 archive as a DMG that happens to be Bzip2-
compressed. Detection of DMG and a handful of other types rely on
finding data partway through or near the ned of a file before
reassigning the entire file as the new type.
Other fixes and considerations in this commit:
- The utf16 HTML parser has weak error handling, particularly with respect
to creating a nested fmap for scanning the ascii decoded file.
This commit cleans up the error handling and wraps the nested scan with
the recursion-stack push()/pop() for correct recursion tracking.
Before this commit, each container layer had a flag to indicate if the
container layer is valid.
We need something similar so that the cli_recursion_stack_get_*()
functions ignore normalized layers. Details...
Imagine an LDB signature for HTML content that specifies a ZIP
container. If the signature actually alerts on the normalized HTML and
you don't ignore normalized layers for the container check, it will
appear as though the alert is in an HTML container rather than a ZIP
container.
This commit accomplishes this with a boolean you set in the scan context
before scanning a new layer. Then when the new fmap is created, it will
use that flag to set similar flag for the layer. The context flag is
reset those that anything after this doesn't have that flag.
The flag allows the new recursion_stack_get() function to ignore
normalized layers when iterating the stack to return a layer at a
requested index, negative or positive.
Scanning normalized extracted/normalized javascript and VBA should also
use the 'layer is normalized' flag.
- This commit also fixes Heuristic.Broken.Executable alert for ELF files
to make sure that:
A) these only alert if cli_append_virus() returns CL_VIRUS (aka it
respects the FP check).
B) all broken-executable alerts for ELF only happen if the
SCAN_HEURISTIC_BROKEN option is enabled.
- This commit also cleans up the error handling in cli_magic_scan_dir().
This was needed so we could correctly apply the layer-is-normalized-flag
to all VBA macros extracted to a directory when scanning the directory.
- Also fix an issue where exceeding scan maximums wouldn't cause embedded
file detection scans to abort. Granted we don't actually want to abort
if max filesize or max recursion depth are exceeded... only if max
scansize, max files, and max scantime are exceeded.
Add 'abort_scan' flag to scan context, to protect against depending on
correct error propagation for fatal conditions. Instead, setting this
flag in the scan context should guarantee that a fatal condition deep in
scan recursion isn't lost which result in more stuff being scanned
instead of aborting. This shouldn't be necessary, but some status codes
like CL_ETIMEOUT never used to be fatal and it's easier to do this than
to verify every parser only returns CL_ETIMEOUT and other "fatal
status codes" in fatal conditions.
- Remove duplicate is_tar() prototype from filestypes.c and include
is_tar.h instead.
- Presently we create the fmap hash when creating the fmap.
This wastes a bit of CPU if the hash is never needed.
Now that we're creating fmap's for all embedded files discovered with
file type recognition scans, this is a much more frequent occurence and
really slows things down.
This commit fixes the issue by only creating fmap hashes as needed.
This should not only resolve the perfomance impact of creating fmap's
for all embedded files, but also should improve performance in general.
- Add allmatch check to the zip parser after the central-header meta
match. That way we don't multiple alerts with the same match except in
allmatch mode. Clean up error handling in the zip parser a tiny bit.
- Fixes to ensure that the scan limits such as scansize, filesize,
recursion depth, # of embedded files, and scantime are always reported
if AlertExceedsMax (--alert-exceeds-max) is enabled.
- Fixed an issue where non-fatal alerts for exceeding scan maximums may
mask signature matches later on. I changed it so these alerts use the
"possibly unwanted" alert-type and thus only alert if no other alerts
were found or if all-match or heuristic-precedence are enabled.
- Added the "Heuristics.Limits.Exceeded.*" events to the JSON metadata
when the --gen-json feature is enabled. These will show up once under
"ParseErrors" the first time a limit is exceeded. In the present
implementation, only one limits-exceeded events will be added, so as to
prevent a malicious or malformed sample from filling the JSON buffer
with millions of events and using a tonne of RAM.
Added a new scan option to alert on broken media (graphics) file
formats. This feature mitigates the risk of malformed media files
intended to exploit vulnerabilities in other software. At present
media validation exists for JPEG, TIFF, PNG, and GIF files.
To enable this feature, set `AlertBrokenMedia yes` in clamd.conf, or
use the `--alert-broken-media` option when using `clamscan`.
These options are disabled by default for now.
Application developers may enable this scan option by enabling
`CL_SCAN_HEURISTIC_BROKEN_MEDIA` for the `heuristic` scan option bit
field.
Fixed PNG parser logic bugs that caused an excess of parsing errors
and fixed a stack exhaustion issue affecting some systems when
scanning PNG files. PNG file type detection was disabled via
signature database update for 0.103.0 to mitigate effects from these
bugs.
Fixed an issue where PNG and GIF files no longer work with Target:5
(graphics) signatures if detected as CL_TYPE_PNG/GIF rather than as
CL_TYPE_GRAPHICS. Target types now support up to 10 possible file
types to make way for additional graphics types in future releases.
Scanning JPEG, TIFF, PNG, and GIF files will no longer return "parse"
errors when file format validation fails. Instead, the scan will alert
with the "Heuristics.Broken.Media" signature prefix and a descriptive
suffix to indicate the issue, provided that the "alert broken media"
feature is enabled.
GIF format validation will no longer fail if the GIF image is missing
the trailer byte, as this appears to be a relatively common issue in
otherwise functional GIF files.
Added a TIFF dynamic configuration (DCONF) option, which was missing.
This will allow us to disable TIFF format validation via signature
database update in the event that it proves to be problematic.
This feature already exists for many other file types.
Added CL_TYPE_JPEG and CL_TYPE_TIFF types.
Many of the core scanning functions' names no longer represent their
specific purpose or arguments. This commit aims to make the names more
intuitive. Names are now prefixed with "magic" if they involve
file-typing and file-type parsing. In addition, each function now
includes the type of input being scanned whether its "desc", "fmap", or
"buff". Some of the APIs also now specify "type" to indicate that a type
other than "ANY" may be passed in to select the type rather than use
file type magic for type recognition.
| current name | new name |
| ------------------------- | --------------------------------- |
| magic_scandesc() | cli_magic_scan() |
| cli_magic_scandesc_type() | <delete> |
| cli_magic_scandesc() | cli_magic_scan_desc() |
| cli_base_scandesc() | cli_magic_scan_desc_type() |
| cli_partition_scandesc() | <delete> |
| cli_map_scandesc() | magic_scan_nested_fmap_type() |
| cli_map_scan() | cli_magic_scan_nested_fmap_type() |
| cli_mem_scandesc() | cli_magic_scan_buff() |
| cli_scanbuff() | cli_scan_buff() |
| cli_scandesc() | cli_scan_desc() |
| cli_fmap_scandesc() | cli_scan_fmap() |
| cli_scanfile() | cli_magic_scan_file() |
| cli_scandir() | cli_magic_scan_dir() |
| cli_filetype2() | cli_determine_fmap_type() |
| cli_filetype() | cli_compare_ftm_file() |
| cli_partitiontype() | cli_compare_ftm_partition() |
| cli_scanraw() | scanraw() |
A way is needed to record scanned file names for two purposes:
1. File names (and extensions) must be stored in the json metadata
properties recorded when using the --gen-json clamscan option. Future
work may use this to compare file extensions with detected file types.
2. File names are useful when interpretting tmp directory output when
using the --leave-temps option.
This commit enables file name retention for later use by storing file
names in the fmap header structure, if a file name exists.
To store the names in fmaps, an optional name argument has been added to
any internal scan API's that create fmaps and every call to these APIs
has been modified to pass a file name or NULL if a file name is not
required. The zip and gpt parsers required some modification to record
file names. The NSIS and XAR parsers fail to collect file names at all
and will require future work to support file name extraction.
Also:
- Added recursive extraction to the tmp directory when the
--leave-temps option is enabled. When not enabled, the tmp directory
structure remains flat so as to prevent the likelihood of exceeding
MAX_PATH. The current tmp directory is stored in the scan context.
- Made the cli_scanfile() internal API non-static and added it to
scanners.h so it would be accessible outside of scanners.c in order to
remove code duplication within libmspack.c.
- Added function comments to scanners.h and matcher.h
- Converted a TDB-type macros and LSIG-type macros to enums for improved
type safey.
- Converted more return status variables from `int` to `cl_error_t` for
improved type safety, and corrected ooxml file typing functions so
they use `cli_file_t` exclusively rather than mixing types with
`cl_error_t`.
- Restructured the magic_scandesc() function to use goto's for error
handling and removed the early_ret_from_magicscan() macro and
magic_scandesc_cleanup() function. This makes the code easier to
read and made it easier to add the recursive tmp directory cleanup to
magic_scandesc().
- Corrected zip, egg, rar filename extraction issues.
- Removed use of extra sub-directory layer for zip, egg, and rar file
extraction. For Zip, this also involved changing the extracted
filenames to be randomly generated rather than using the "zip.###"
file name scheme.
EGG extraction support includes deflate, bzip2, and lzma decompression. AZO (LZO?) decompression not yet supported. Solid archives not yet supported. Split archives may have some limited success.
This commit also includes updates to autoconf iconv.m4 file enable detection of libiconv in alternative install locations.
Val Snyder