vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach
vLLM is an inference and serving engine for large language models. From 0.24.0 until 0.30.0, the Qwen2VLVideoBackend and Qwen3VLVideoBackend classes accept request-level values for the media_io_kwargs.video.max_frames and media_io_kwargs.video.fps fields without enforcing server-side ceilings. An unauthenticated caller can submit these values to the /tokenize endpoint, causing the sampler to decode every frame selected from attacker-controlled video input, consume disproportionate frontend memory, and potentially terminate the API process before scheduling or admission control. The Rust frontend is not affected because it rejects the media_io
vLLM: Structured-output request errors escape the request boundary and terminate the shared EngineCore — engine-fatal denial of service (3 sites)
vLLM is an inference and serving engine for large language models. Prior to 0.30.0, structured-output request failures can escape request-scoped validation and reach the EngineCore fatal-error path. A per-request backend mismatch can re-raise a grammar compilation exception, padding produced by the ngram_gpu speculative-decoding mode can pass a negative token to guidance validation, and the Rust frontend can admit empty structured-output values that the Python frontend rejects, allowing ordinary constrained-generation requests to terminate the shared engine. This issue is fixed in version 0.30.0.
vLLM: Loose `cache_salt` validation lets a single request kill EngineCore on LMCache-MP deployments — uncaught downstream `ValueError` denial of service
vLLM is an inference and serving engine for large language models. Prior to 0.30.0, OpenAI-compatible request models accept a non-empty cache_salt value without enforcing the character and length restrictions required by the IPCCacheServerKey consumer in LMCache-MP. On deployments using the LMCache-MP connector, a salt that contains a forbidden character or exceeds the permitted length can raise an uncaught ValueError during scheduler cache lookup, causing EngineCore to terminate and denying service to all concurrent users. This issue is fixed in version 0.30.0.
Quart leaks raw request body (incl. plaintext passwords) to stdout via stray debug print in Body.__await__
### Summary
Quart 0.23.0 contains a stray debug statement (`print(data)`) inside `Body.__await__` in `quart/wrappers/request.py`. Any request whose body is awaited — `await request.form`, `await request.get_data()`, WTForms `validate_on_submit()`, etc. — has its raw, unparsed body printed to stdout, including plaintext form fields such as passwords and CSRF tokens. Confirmed present in 0.23.0, confirmed absent in 0.22.0.
### Details
In `src/quart/wrappers/request.py`, `Body.__await__` accumulates the request body into a bytearray:
```python
data = bytearray()
while not self._queue.empty():
data.extend(self._queue.get_nowait())
prin
vLLM: Flash late-interaction scoring caches query embeddings under a caller-controlled request id — cross-request integrity break and induced errors on `/score` and `/rerank`
vLLM is an inference and serving engine for large language models. Prior to 0.30.0, flash late-interaction scoring at the /score and /rerank endpoints derives each worker's query_key value from the caller-controlled X-Request-Id header. A concurrent request that reuses a victim's identifier can overwrite the cached query embedding so the victim's documents are scored against the attacker's query, and shared use counters can also cause a late-interaction cache-miss error. This issue is fixed in version 0.30.0.
vLLM: Scale-out disaggregated multimodal transport trusts caller-supplied features
vLLM is an inference and serving engine for large language models. Prior to 0.30.0, the /inference/v1/generate endpoint in the disaggregated scale-out path accepts caller-supplied tensors in the features.kwargs_data field, cache identifiers in the features.mm_hashes field, ranges in the features.mm_placeholders field, and wire-selected multimodal field processors without rebinding them to the active model renderer contract. Forged grid geometry, field types, or non-positive placeholder lengths can terminate the shared EngineCore; when an attacker knows or can induce a victim's content hash, forged cache hashes can poison or retrieve cross-req
vLLM: Mirrored multimodal IPC caches desync after a rejected request — a later request reusing the same media hash trips a receiver assertion in the engine core
vLLM is an inference and serving engine for large language models. Prior to 0.28.0, the default mirrored multimodal LRU cache can commit a media hash in the frontend sender cache during multimodal rendering and before engine admission, while the engine receiver cache never receives the payload if that request is rejected. A later request reusing the same media hash causes MultiModalProcessorSenderCache to send no payload and MultiModalReceiverCache to reach an assertion with the message "Expected a cached item," producing a shared-service availability failure. This issue is fixed in version 0.28.0.
Low · CVSS 3.1EPSS 0.22%Public exploitFix available
vLLM: Harmony tool continuations drop `cache_salt` — restoring a cross-tenant prefix-cache membership oracle
vLLM is an inference and serving engine for large language models. Prior to 0.30.0, Harmony tool continuations submitted through "POST /v1/responses" requests rebuild the next-turn engine input without preserving the cache_salt value, placing the continuation prefix in the global unsalted cache namespace even when the caller enabled salting. On deployments with prefix caching enabled, which is the default, an authenticated tenant who can reconstruct a victim's low-entropy post-tool history can submit the same continuation and use the cached_tokens_per_turn count to determine whether the prefix was previously processed, defeating the intended
Docling: Arbitrary local file read via draw:image xlink:href in the OpenDocument backend
Docling simplifies document processing by parsing diverse formats and providing integrations with the generative AI ecosystem. From 2.107.0 until 2.120.3, docling/backend/opendocument_backend.py uses the xlink:href attribute value of a draw:image element as a filesystem path when the referenced part is not found in the document archive. The _image_ref_from_odf_image function reads that attacker-controlled path without a scheme check, extraction-directory confinement, or the enable_local_fetch setting used by other backends. Readable files that Pillow can decode as images are embedded in converted output, and other existing paths can be distin
Docling: `enable_local_fetch` is not enforced in HTML browser-rendering mode
Docling simplifies document processing by parsing diverse formats and providing integrations with the generative AI ecosystem. From 2.82.0 until 2.118.1, HTMLBackendOptions(render_page=True) permits file URLs because HTMLDocumentBackend._get_browser_request_block_reason does not enforce the enable_local_fetch setting or confine local requests to the source document directory. Crafted path-backed HTML can embed a readable local text file in a browser-rendered page image when Playwright is installed. Only filesystem Path inputs are affected because stream inputs use an opaque origin, and the default configuration, command-line interface, doclin
Medium · CVSS 6.5EPSS 0.29%Public exploitFix available
Docling: Unbounded table rowspan/colspan in HTML, JATS, ODS and BoxNote backends causes CPU/memory exhaustion
Docling simplifies document processing by parsing diverse formats and providing integrations with the generative AI ecosystem. From 2.0.0 until 2.131.0, the HTML, JATS, OpenDocument spreadsheet, and BoxNote backends, including docling/backend/html_backend.py, docling/backend/jats_backend.py, and docling/backend/boxnote_backend.py, accept the rowspan and colspan attribute values without an upper bound and execute loops or allocate a table grid proportional to the declared span. A very small document can therefore cause sustained CPU use or multi-gigabyte memory allocation, and the document_timeout setting does not interrupt the single backend
Medium · CVSS 4.3EPSS 0.3%Public exploitFix available
Docling: Crafted DoclingDocument JSON embeds local image files into converted output
Docling simplifies document processing by parsing diverse formats and providing integrations with the generative AI ecosystem. From 2.16.0 until 2.131.0, the InputFormat.JSON_DOCLING backend in docling/backend/json/docling_json_backend.py validates serialized DoclingDocument input without rejecting picture image references that contain local paths or file URIs. When the document is enriched or exported with ImageRefMode.EMBEDDED, the DoclingDocument._with_embedded_pictures and ImageRef.pil_image methods can open those references and place readable image bytes in Markdown or HTML output. Disclosure is limited to files Pillow can decode as imag
Docling: METS-GBS archive member limit enforced after full member enumeration (memory exhaustion during format detection)
Docling simplifies document processing by parsing diverse formats and providing integrations with the generative AI ecosystem. From 2.45.0 until 2.131.0, METS-GBS format detection in docling/datamodel/document.py and the backend in docling/backend/mets_gbs_backend.py call tarfile.TarFile.getmembers() before enforcing the max_member_count limit, causing the full archive member list to be allocated before the limit can stop processing. A small gzip-compressed tar archive with a very large number of empty members can therefore consume memory proportional to the declared member count, including during format detection before the allowed_formats r
Docling: KServe v2 OCR engine does not enforce enable_remote_services
Docling simplifies document processing by parsing diverse formats and providing integrations with the generative AI ecosystem. From 2.83.0 until 2.131.0, the KServeV2OcrModel class defined in docling/models/stages/ocr/kserve_v2_ocr_model.py sends page images to its configured endpoint without checking the pipeline_options.enable_remote_services setting, even when the caller sets that policy control to false. The StandardPdfPipeline._make_ocr_model method also fails to pass the flag into the OCR factory, allowing remote OCR processing in configurations that rely on remote services being disabled. The destination is configured by the caller rat
Docling: Plugin entry points are imported before the allow_external_plugins check
Docling simplifies document processing by parsing diverse formats and providing integrations with the generative AI ecosystem. From 2.27.0 until 2.131.0, Docling plugin factories in docling/models/factories/base_factory.py call load_setuptools_entrypoints() before applying the allow_external_plugins setting, so every module registered in the Docling entry-point group is imported even when external plugins are disabled. An installed third-party or compromised package can therefore execute import-time code when Docling starts, while the subsequent namespace filter misleadingly reports that the plugin was not loaded. This issue is fixed in 2.131
High · CVSS 7.5EPSS 0.37%Public exploitFix available
Docling: Arbitrary file read/write (and command execution when shell-escape is enabled) when rendering untrusted TikZ with the opt-in Tectonic engine
Docling simplifies document processing by parsing diverse formats and providing integrations with the generative AI ecosystem. From 2.94.0 until 2.132.0, callers that opt into LatexBackendOptions(tikz_engine="tectonic") invoke docling/backend/latex/engines/tectonic.py to compile an untrusted TikZ body and document preamble without restricting TeX file primitives including \openin and \openout. Crafted input can read files available to the converter and create or overwrite writable files, and enabling the tikz_engine_allow_shell_escape option additionally permits shell commands through TeX. The default configuration, which does not enable Tect
Docling: SSRF guard bypass in remote resource fetching (DNS rebinding / multi-record resolution; no IP validation in HTML render mode)
Docling simplifies document processing by parsing diverse formats and providing integrations with the generative AI ecosystem. From 2.91.0 until 2.132.0, validate_url_safety in docling/backend/utils/image_resource_loader.py validates a hostname with a single IPv4 lookup and then allows the HTTP client to resolve and parse the original URL again, permitting DNS rebinding, mixed public and internal address records, and backslash authority parser disagreement to reach internal services. HTMLBackendOptions(render_page=True) also allows HTTP and HTTPS browser requests without validating their resolved destination. Exploitation requires remote fetc
Low · CVSS 3.7EPSS 0.25%Public exploitFix available
Docling: Configured HTTP headers sent to every remote image host named by a document
Docling simplifies document processing by parsing diverse formats and providing integrations with the generative AI ecosystem. From 2.95.0 until 2.132.0, the HTML image resource loader in docling/backend/utils/image_resource_loader.py forwards headers configured through the HTMLBackendOptions.headers setting to every remote image URL named by an untrusted document when enable_remote_fetch=True and fetch_images=True. The loader does not restrict those credentials to the source document's origin, allowing requests that carry custom headers such as API keys and cookies to follow cross-origin redirects and expose the caller's configured credentia
Langflow: IP Spoofing Bypass via `X-Forwarded-For` Allowing Remote Configuration Write
Langflow is a tool for building and deploying AI-powered agents and workflows. From 1.5.0 until 1.10.3, an IP spoofing vulnerability in the Model Context Protocol (MCP) configuration installation endpoint (POST /api/v1/mcp/project/{project_id}/install) allowed authenticated remote attackers to bypass the "local-only" access restriction. By sending a spoofed X-Forwarded-For: 127.0.0.1 header, an attacker could make the server treat the request as originating from localhost, letting them write/overwrite an MCP client configuration file on the server's filesystem. This vulnerability is fixed in 1.10.3.
Langflow: Authenticated Cross-Project File Disclosure via Unscoped MCP Resource Handlers
Langflow is a tool for building and deploying AI-powered agents and workflows. From 1.6.8 until 1.9.1, Langflow authenticated access to the project identifier in a project-scoped MCP connection but did not authorize the resource URI supplied to resources/read. read_resource forwarded the attacker-controlled URI to handle_read_resource, which parsed a flow_id and filename and called storage_service.get_file without verifying that the flow belonged to the authenticated user or current project. A user with access to any project-scoped MCP endpoint could therefore request another user's flow-backed file, while global handle_list_resources and han