GitLab Advisory Database (GLAD)

vLLM's default multimodal cache (mm_processor_cache_type="lru") mirrors state across two processes: the frontend (P0) holds only metadata (MultiModalProcessorSenderCache) while the engine core (P1) holds the real payload (MultiModalReceiverCache). The design invariant is that get_and_update() runs on P0 and P1 in lockstep for every request, so eviction order stays mirrored and P0 can answer "is this cached in P1?" without talking to P1. That invariant breaks when a request is rejected after …

On the GPT-OSS "Harmony" path (POST /v1/responses), a request that uses a built-in or MCP tool runs as a multi-turn loop: after each tool call vLLM re-renders the full next-turn Harmony prompt and re-submits it to the engine. Turn 1 correctly carries request.cache_salt, but the tool-continuation re-submission rebuilds the engine input via tokens_input(token_ids) with no cache_salt. The continuation prefix is therefore cached in the global unsalted namespace even though the …

When the HTML backend renders pages in a headless browser (HTMLBackendOptions(render_page=True)), the enable_local_fetch option is not enforced. A crafted HTML file can embed an arbitrary local file (for example with <iframe src="file:///…">), and that file's contents appear in the page image attached to the returned DoclingDocument.

When the HTML backend renders pages in a headless browser (HTMLBackendOptions(render_page=True)), the enable_local_fetch option is not enforced. A crafted HTML file can embed an arbitrary local file (for example with <iframe src="file:///…">), and that file's contents appear in the page image attached to the returned DoclingDocument.

Werkzeug's safe_join function allows Windows device names as filenames when they have an empty ADS marker on NTFS. This was previously reported as https://github.com/pallets/werkzeug/security/advisories/GHSA-hgf8-39gv-g3f2, but the added filtering failed to account for the fact Windows allows special device names with an empty ADS marker, such as NUL:. send_from_directory uses safe_join to safely serve files at user-specified paths under a directory. If the application is running on Windows and NTFS, and …

NodeVM exposes the host util module to the sandbox through an unfiltered shallow copy (Object.assign({}, util)). On Node.js >= 22.9 this hands sandboxed code util.getCallSites(), a programmatic stack-introspection API that returns the host process's full call stack — absolute file paths, function names, and line numbers — including vm2 bridge internals and the embedding application's entrypoint. This bypasses the host-frame redaction established in GHSA-v27g-jcqj-v8rw, which only covers the Error.prepareStackTrace channel.

vm2's VM({ timeout }) option is documented and relied upon as the mechanism that bounds how long sandboxed code may execute. In the current implementation, the timeout only wraps the single synchronous call to VM#run() (via doWithTimeout → this._runScript(script) in lib/vm.js). It does not, and structurally cannot, bound code that the V8 engine itself schedules to run after that call has already returned. FinalizationRegistry and WeakRef are exposed to sandboxed …

Sandboxed code is able to disclose host memory used by small allocations by Buffer.from, Buffer.concat.

Insufficient clearing of the output buffer in Java-based decompressor implementations in lz4-java 1.10.0 and earlier allows remote attackers to read previous buffer contents via crafted compressed input. In applications where the output buffer is reused without being cleared, this may lead to disclosure of sensitive data. JNI-based implementations are not affected.

Insufficient clearing of the output buffer in Java-based decompressor implementations in lz4-java 1.10.0 and earlier allows remote attackers to read previous buffer contents via crafted compressed input. In applications where the output buffer is reused without being cleared, this may lead to disclosure of sensitive data. JNI-based implementations are not affected.

Insufficient clearing of the output buffer in Java-based decompressor implementations in lz4-java 1.10.0 and earlier allows remote attackers to read previous buffer contents via crafted compressed input. In applications where the output buffer is reused without being cleared, this may lead to disclosure of sensitive data. JNI-based implementations are not affected.

This advisory has been marked as a false positive.

This advisory has been marked as a false positive.

This advisory has been marked as a false positive.

This advisory has been marked as False Positive and has been removed.

After an API Token exposure from an exploited trivy dependency, two new releases of litellm were uploaded to PyPI containing automatically activated malware, harvesting sensitive credentials and files, and exfiltrating to a remote API. Anyone who has installed and run the project should assume any credentials available to litellm environment may have been exposed, and revoke/rotate thema ccordingly.