Vulnerabilities

vLLM is an inference and serving engine for large language models (LLMs). Prior to version 0.14.1, a Server-Side Request Forgery (SSRF) vulnerability exists in the `MediaConnector` class within the vLLM project's multimodal feature set. The load_from_url and load_from_url_async methods obtain and process media from URLs provided by users, using different Python parsing libraries when restricting the target host. These two parsing libraries have different interpretations of backslashes, which allows the host name restriction to be bypassed. This allows an attacker to coerce the vLLM server into making arbitrary requests to internal network resources. This vulnerability is particularly critical in containerized environments like `llm-d`, where a compromised vLLM pod could be used to scan the internal network, interact with other pods, and potentially cause denial of service or access sensitive data. For example, an attacker could make the vLLM pod send malicious requests to an internal `llm-d` management endpoint, leading to system instability by falsely reporting metrics like the KV cache state. Version 0.14.1 contains a patch for the issue.

soroban-fixed-point-math is a fixed-point math library for Soroban smart contacts. In versions 1.3.0 and 1.4.0, the `mulDiv(x, y, z)` function incorrectly handled cases where both the intermediate product $x * y$ and the divisor $z$ were negative. The logic assumed that if the intermediate product was negative, the final result must also be negative, neglecting the sign of $z$. This resulted in rounding being applied in the wrong direction for cases where both $x * y$ and $z$ were negative. The functions most at risk are `fixed_div_floor` and `fixed_div_ceil`, as they often use non-constant numbers as the divisor $z$ in `mulDiv`. This error is present in all signed `FixedPoint` and `SorobanFixedPoint` implementations, including `i64`, `i128`, and `I256`. Versions 1.3.1 and 1.4.1 contain a patch. No known workarounds for this issue are available.

ConvertXis a self-hosted online file converter. In versions prior to 0.17.0, the `POST /delete` endpoint uses a user-controlled `filename` value to construct a filesystem path and deletes it via `unlink` without sufficient validation. By supplying path traversal sequences (e.g., `../`), an attacker can delete arbitrary files outside the intended uploads directory, limited only by the permissions of the server process. Version 0.17.0 fixes the issue.

PyTorch is a Python package that provides tensor computation. Prior to version 2.10.0, a vulnerability in PyTorch's `weights_only` unpickler allows an attacker to craft a malicious checkpoint file (`.pth`) that, when loaded with `torch.load(..., weights_only=True)`, can corrupt memory and potentially lead to arbitrary code execution. Version 2.10.0 fixes the issue.

Kargo manages and automates the promotion of software artifacts. Prior to versions 1.8.7, 1.7.7, and 1.6.3, a bug was found with authentication checks on the `GetConfig()` API endpoint. This allowed unauthenticated users to access this endpoint by specifying an `Authorization` header with any non-empty `Bearer` token value, regardless of validity. This vulnerability did allow for exfiltration of configuration data such as endpoints for connected Argo CD clusters. This data could allow an attacker to enumerate cluster URLs and namespaces for use in subsequent attacks. Additionally, the same bug affected the `RefreshResource` endpoint. This endpoint does not lead to any information disclosure, but could be used by an unauthenticated attacker to perform a denial-of-service style attack against the Kargo API. `RefreshResource` sets an annotation on specific Kubernetes resources to trigger reconciliations. If run on a constant loop, this could also slow down legitimate requests to the Kubernetes API server. This problem has been patched in Kargo versiosn 1.8.7, 1.7.7, and 1.6.3. There are no workarounds for this issue.

PHPUnit is a testing framework for PHP. A vulnerability has been discovered in versions prior to 12.5.8, 11.5.50, 10.5.62, 9.6.33, and 8.5.52 involving unsafe deserialization of code coverage data in PHPT test execution. The vulnerability exists in the `cleanupForCoverage()` method, which deserializes code coverage files without validation, potentially allowing remote code execution if malicious `.coverage` files are present prior to the execution of the PHPT test. The vulnerability occurs when a `.coverage` file, which should not exist before test execution, is deserialized without the `allowed_classes` parameter restriction. An attacker with local file write access can place a malicious serialized object with a `__wakeup()` method into the file system, leading to arbitrary code execution during test runs with code coverage instrumentation enabled. This vulnerability requires local file write access to the location where PHPUnit stores or expects code coverage files for PHPT tests. This can occur through CI/CD pipeline attacks, the local development environment, and/or compromised dependencies. Rather than just silently sanitizing the input via `['allowed_classes' => false]`, the maintainer has chosen to make the anomalous state explicit by treating pre-existing `.coverage` files for PHPT tests as an error condition. Starting in versions in versions 12.5.8, 11.5.50, 10.5.62, 9.6.33, when a `.coverage` file is detected for a PHPT test prior to execution, PHPUnit will emit a clear error message identifying the anomalous state. Organizations can reduce the effective risk of this vulnerability through proper CI/CD configuration, including ephemeral runners, code review enforcement, branch protection, artifact isolation, and access control.

RAGFlow is an open-source RAG (Retrieval-Augmented Generation) engine. In version 0.23.1 and possibly earlier versions, the MinerU parser contains a "Zip Slip" vulnerability, allowing an attacker to overwrite arbitrary files on the server (leading to Remote Code Execution) via a malicious ZIP archive. The MinerUParser class retrieves and extracts ZIP files from an external source (mineru_server_url). The extraction logic in `_extract_zip_no_root` fails to sanitize filenames within the ZIP archive. Commit 64c75d558e4a17a4a48953b4c201526431d8338f contains a patch for the issue.

gmrtd is a Go library for reading Machine Readable Travel Documents (MRTDs). Prior to version 0.17.2, ReadFile accepts TLVs with lengths that can range up to 4GB, which can cause unconstrained resource consumption in both memory and cpu cycles. ReadFile can consume an extended TLV with lengths well outside what would be available in ICs. It can accept something all the way up to 4GB which would take too many iterations in 256 byte chunks, and would also try to allocate memory that might not be available in constrained environments like phones. Or if an API sends data to ReadFile, the same problem applies. The very small chunked read also locks the goroutine in accepting data for a very large number of iterations. projects using the gmrtd library to read files from NFCs can experience extreme slowdowns or memory consumption. A malicious NFC can just behave like the mock transceiver described above and by just sending dummy bytes as each chunk to be read, can make the receiving thread unresponsive and fill up memory on the host system. Version 0.17.2 patches the issue.

Dozzle is a realtime log viewer for docker containers. Prior to version 9.0.3, a flaw in Dozzle’s agent-backed shell endpoints allows a user restricted by label filters (for example, `label=env=dev`) to obtain an interactive root shell in out‑of‑scope containers (for example, `env=prod`) on the same agent host by directly targeting their container IDs. Version 9.0.3 contains a patch for the issue.

Squidex is an open source headless content management system and content management hub. Versions of the application up to and including 7.21.0 allow users to define "Webhooks" as actions within the Rules engine. The url parameter in the webhook configuration does not appear to validate or restrict destination IP addresses. It accepts local addresses such as 127.0.0.1 or localhost. When a rule is triggered (Either manual trigger by manually calling the trigger endpoint or by a content update or any other triggers), the backend server executes an HTTP request to the user-supplied URL. Crucially, the server logs the full HTTP response in the rule execution log (lastDump field), which is accessible via the API. Which turns a "Blind" SSRF into a "Full Read" SSRF. As of time of publication, no patched versions are available.

Inappropriate implementation in Background Fetch API in Google Chrome prior to 144.0.7559.110 allowed a remote attacker to leak cross-origin data via a crafted HTML page. (Chromium security severity: High)

An Authentication Bypass Using an<br /> Alternate Path or Channel vulnerability in Juniper Networks Session Smart<br /> Router may allows a network-based attacker to bypass authentication<br /> and take administrative control of the device.<br /> <br /> This issue affects Session Smart Router: <br /> <br /> <br /> <br /> * from 5.6.7 before 5.6.17, <br /> * from 6.0 before 6.0.8 (affected from 6.0.8),<br /> <br /> * from 6.1 before 6.1.12-lts, <br /> * from 6.2 before 6.2.8-lts, <br /> * from 6.3 before 6.3.3-r2; <br /> <br /> <br /> <br /> <br /> This issue affects Session Smart Conductor: <br /> <br /> <br /> <br /> * from 5.6.7 before 5.6.17, <br /> * from 6.0 before 6.0.8 (affected from 6.0.8),<br /> <br /> * from 6.1 before 6.1.12-lts, <br /> * from 6.2 before 6.2.8-lts, <br /> * from 6.3 before 6.3.3-r2; <br /> <br /> <br /> <br /> <br /> This issue affects WAN Assurance Managed Routers: <br /> <br /> <br /> <br /> * from 5.6.7 before 5.6.17, <br /> * from 6.0 before 6.0.8 (affected from 6.0.8),<br /> <br /> * from 6.1 before 6.1.12-lts, <br /> * from 6.2 before 6.2.8-lts, <br /> * from 6.3 before 6.3.3-r2.