A security flaw has been discovered in vibrantlabsai RAGAS up to 0.4.3. The affected element is the function _try_process_local_file/_try_process_url of the file src/ragas/metrics/collections/multi_modal_faithfulness/util.py of the component Collections Module. Performing a manipulation of the argument retrieved_contexts results in server-side request forgery. The attack can be initiated remotely. The exploit has been released to the public and may be used for attacks. The security patch for CVE-2025-45691 was applied to a different module only. The vendor was contacted early about this disclosure but did not respond in any way.

Vim is an open source, command line text editor. Prior to version 9.2.0597, Vim's Python omni-completion executes reconstructed function and class definitions from the current buffer with exec() as part of populating the completion dictionary. Python evaluates function default values, parameter annotations, and class base expressions at definition time, so a hostile buffer can execute attacker-controlled Python expressions during omni-completion. The existing g:pythoncomplete_allow_import mitigation (GHSA-52mc-rq6p-rc7c) does not cover this path, because the attacker-controlled code is not a harvested import/from statement. This issue has been patched in version 9.2.0597.

A flaw was found in Keycloak. An authenticated attacker can perform Server-Side Request Forgery (SSRF) by manipulating the `client_session_host` parameter during refresh token requests. This occurs when a Keycloak client is configured to use the `backchannel.logout.url` with the `application.session.host` placeholder. Successful exploitation allows the attacker to make HTTP requests from the Keycloak server’s network context, potentially probing internal networks or internal APIs, leading to information disclosure.

Starlette is a lightweight ASGI framework/toolkit. In versions 1.0.1 and earlier, StaticFiles on Windows is vulnerable to SSRF. An UNC path such as \\attacker.com\share can cause os.path.realpath to initiate an outbound SMB connection before the path is rejected, exposing the service account’s NTLMv2 credentials for offline cracking or relay even though the HTTP response is only a 404. The issue affects default follow_symlink=False deployments, including frameworks built on Starlette such as FastAPI; POSIX systems and follow_symlink=True are unaffected. The issue is fixed in 1.1.0.

IO::Compress versions before 2.220 for Perl can execute arbitrary code in File::GlobMapper via an attacker-controlled output glob. _parseOutputGlob() wraps the caller-supplied output glob string in double quotes and stores it in the parser state; _getFiles() then runs the stored expression through eval STRING. A literal double quote in the output glob closes the dquote wrapper, and the characters that follow are evaluated as Perl. Arbitrary Perl in the output glob executes at the calling process's privilege.

Vim is an open source, command line text editor. Prior to version 9.2.0495, a Vimscript code injection vulnerability exists in s:NetrwBookHistSave() in the netrw plugin (runtime/pack/dist/opt/netrw/autoload/netrw.vim) when serializing browsed directory paths to the history file ~/.vim/.netrwhist. A directory name derived from the filesystem is interpolated into a single-quoted Vimscript string literal without escaping embedded single quotes, allowing a crafted directory name to break out of the string context and execute arbitrary Vimscript, including shell commands via system() and :!, the next time the history file is sourced. This issue has been patched in version 9.2.0495.

Versions of the package simple-git before 3.36.0 are vulnerable to Remote Code Execution (RCE) due to an incomplete fix for [CVE-2022-25912](https://security.snyk.io/vuln/SNYK-JS-SIMPLEGIT-3112221) that blocks the -c option but not the equivalent --config form. If untrusted input can reach the options argument passed to simple-git, an attacker may still achieve remote code execution by enabling protocol.ext.allow=always and using an ext:: clone source.

A pre-authentication, code injection vulnerability in version 1.0.0 or later of the ChromaDB Python project allows an unauthenticated attacker to run arbitrary code on the server by sending a malicious model repository and trust_remote_code set to true in the /api/v2/tenants/{tenant}/databases/{db}/collections endpoint.

A code injection vulnerability in version 0.4.17 or later of the ChromaDB Python project allows an authenticated attacker to run arbitrary code on the server by sending a malicious model repository and trust_remote_code set to true in the /api/v2/tenants/default_tenant/databases/default_database/collections/{collection_id} if they have the UPDATE_COLLECTION permission.

protobufjs compiles protobuf definitions into JavaScript (JS) functions. Prior to 7.5.6 and 8.0.2, protobufjs generated JavaScript for toObject conversion could include an unsafe expression derived from a schema-controlled bytes field default value. A crafted descriptor with a non-string default value for a bytes field could cause attacker-controlled code to be emitted into the generated conversion function. This vulnerability is fixed in 7.5.6 and 8.0.2.

vm2 is an open source vm/sandbox for Node.js. From 3.9.6 to 3.10.5, vm2's bridge exposes mutable proxies for real host-realm intrinsic prototypes and then forwards sandbox writes into the underlying host objects with otherReflectSet() and otherReflectDefineProperty(), which lets attacker-controlled JavaScript running in a default VM or inherited NodeVM mutate shared host Object.prototype, Array.prototype, and Function.prototype from inside the sandbox This vulnerability is fixed in 3.11.0.

DiskCache (python-diskcache) through 5.6.3 uses Python pickle for serialization by default. An attacker with write access to the cache directory can achieve arbitrary code execution when a victim application reads from the cache.

Axios is a promise based HTTP client for the browser and Node.js. From 0.19.0 to before 0.31.1 and 1.15.2, Axios contains prototype-pollution gadgets in request config processing. If another vulnerability in the same JavaScript process has already polluted Object.prototype.transformResponse, affected Axios versions may treat that inherited value as request configuration or as an option validator. Axios does not itself create the prototype pollution. Exploitability requires a separate prototype-pollution vulnerability or equivalent attacker control over Object.prototype before Axios creates a request. This vulnerability is fixed in 0.31.1 and 1.15.2.

A flaw was identified in Keycloak, an identity and access management solution, where it improperly follows HTTP redirects when processing certain client configuration requests. This behavior allows an attacker to trick the server into making unintended requests to internal or restricted resources. As a result, sensitive internal services such as cloud metadata endpoints could be accessed. This issue may lead to information disclosure and enable attackers to map internal network infrastructure.

Impact: The fix for CVE-2021-23337 (https://github.com/advisories/GHSA-35jh-r3h4-6jhm) added validation for the variable option in _.template but did not apply the same validation to options.imports key names. Both paths flow into the same Function() constructor sink. When an application passes untrusted input as options.imports key names, an attacker can inject default-parameter expressions that execute arbitrary code at template compilation time. Additionally, _.template uses assignInWith to merge imports, which enumerates inherited properties via for..in. If Object.prototype has been polluted by any other vector, the polluted keys are copied into the imports object and passed to Function(). Patches: Users should upgrade to version 4.18.0. Workarounds: Do not pass untrusted input as key names in options.imports. Only use developer-controlled, static key names.

vm2 is an open source vm/sandbox for Node.js. Prior to 3.11.0, It is possible to reach BaseHandler.getPrototypeOf, which can be used to get arbitrary prototypes. This vulnerability is fixed in 3.11.0.

Next.js is a React framework for building full-stack web applications. From 13.4.13 to before 15.5.16 and 16.2.5, self-hosted applications using the built-in Node.js server can be vulnerable to server-side request forgery through crafted WebSocket upgrade requests. An attacker can cause the server to proxy requests to arbitrary internal or external destinations, which may expose internal services or cloud metadata endpoints. Vercel-hosted deployments are not affected. This vulnerability is fixed in 15.5.16 and 16.2.5.

Diffusers is the a library for pretrained diffusion models. Prior to 0.38.0, a trust_remote_code bypass in DiffusionPipeline.from_pretrained allows arbitrary remote code execution despite the user passing trust_remote_code=False (or omitting it, which is the default). The vulnerability has three variants, all sharing the same root cause — the trust_remote_code gate was implemented inside DiffusionPipeline.download() rather than at the actual dynamic-module load site, so any code path that bypassed or short-circuited download() also bypassed the security check. DiffusionPipeline.from_pretrained('repoA', custom_pipeline='attacker/repoB', trust_remote_code=False) — the gate evaluated against repoA's file list rather than repoB's, so repoB's pipeline.py was loaded and executed. DiffusionPipeline.from_pretrained('/local/snapshot', custom_pipeline='attacker/repoB', trust_remote_code=False) — the local-path branch never invoked download(), so the gate was never reached and remote code from repoB executed. DiffusionPipeline.from_pretrained('/local/snapshot', trust_remote_code=False) where the snapshot contains custom component files (e.g. unet/my_unet_model.py) referenced from model_index.json — same root cause; the local path skipped download() and custom component code executed. This vulnerability is fixed in 0.38.0.

Docling-Graph turns documents into validated Pydantic objects, then builds a directed knowledge graph with explicit semantic relationships. Prior to 1.5.1, the URLInputHandler class in docling_graph/core/input/handlers.py makes HTTP requests to user-supplied URLs without validating whether the target resolves to a private, loopback, or link-local IP address. The URLValidator only checks for a valid scheme and non-empty netloc, performing no IP-level validation. Additionally, requests.head() was called with allow_redirects=True, allowing an attacker to redirect requests to internal endpoints via an intermediary URL. An attacker who can control the --source CLI argument or PipelineConfig.source API parameter can trigger Server-Side Request Forgery (SSRF). This vulnerability is fixed in 1.5.1.

Angular is a development platform for building mobile and desktop web applications using TypeScript/JavaScript and other languages. Prior to 22.0.0-next.12, 21.2.13, 20.3.21, and 19.2.22, a Server-Side Request Forgery (SSRF) vulnerability exists in @angular/platform-server. The issue stems from how the server-side rendering (SSR) engine processes the request URL provided to the rendering entry points. When an absolute-form URL (e.g., http://evil.com) is passed to the rendering engine, the internal ServerPlatformLocation can be manipulated into adopting the attacker-controlled domain as the "current" hostname. Consequently, any relative HttpClient requests or PlatformLocation.hostname references are redirected to the attacker controlled server, potentially exposing internal APIs or metadata services. This vulnerability is fixed in 22.0.0-next.12, 21.2.13, 20.3.21, and 19.2.22.

# Active Storage allowed transformation methods potentially unsafe Active Storage attempts to prevent the use of potentially unsafe image transformation methods and parameters by default. The default allowed list contains three methods allow for the circumvention of the safe defaults which enables potential command injection vulnerabilities in cases where arbitrary user supplied input is accepted as valid transformation methods or parameters. Impact ------ This vulnerability impacts applications that use Active Storage with the image_processing processing gem in addition to mini_magick as the image processor. Vulnerable code will look something similar to this: ``` <%= image_tag blob.variant(params[:t] => params[:v]) %> ``` Where the transformation method or its arguments are untrusted arbitrary input. All users running an affected release should either upgrade or use one of the workarounds immediately. Workarounds ----------- Consuming user supplied input for image transformation methods or their parameters is unsupported behavior and should be considered dangerous. Strict validation of user supplied methods and parameters should be performed as well as having a strong [ImageMagick security policy](https://imagemagick.org/script/security-policy.php) deployed. Credits ------- Thank you [lio346](https://hackerone.com/lio346) for reporting this!

vm2 is an open source vm/sandbox for Node.js. Prior to 3.11.0, it is possible to obtain the host Object. There are various ways to use the host Object, to escape the sandbox, one example would be using HostObject.getOwnPropertySymbols to obtain Symbol(nodejs.util.inspect.custom). This vulnerability is fixed in 3.11.0.

A flaw was found in the OpenShift Router. A user with EndpointSlice write access can exploit this vulnerability by creating a Service backed by an FQDN (Fully Qualified Domain Name) EndpointSlice that resolves to a cloud metadata endpoint. This allows the router to proxy requests to the cloud metadata endpoint, leading to the disclosure of instance credentials and other sensitive metadata. This bypasses previous security measures for validating IP addresses.

python jsonpickle 2.0.0 contains a remote code execution vulnerability that allows attackers to execute arbitrary Python commands by deserializing malicious JSON payloads containing py/repr objects. Attackers can craft JSON strings with py/repr directives that invoke the eval function during deserialization to execute arbitrary code.

LiteLLM is a proxy server (AI Gateway) to call LLM APIs in OpenAI (or native) format. From version 1.80.5 to before version 1.83.7, the POST /prompts/test endpoint accepted user-supplied prompt templates and rendered them without sandboxing. A crafted template could run arbitrary code inside the LiteLLM Proxy process. The endpoint only checks that the caller presents a valid proxy API key, so any authenticated user could reach it. Depending on how the proxy is deployed, this could expose secrets in the process environment (such as provider API keys or database credentials) and allow commands to be run on the host. This issue has been patched in version 1.83.7.

Contour is a Kubernetes ingress controller using Envoy proxy. From v1.19.0 to before v1.33.4, v1.32.5, and v1.31.6, Contour's Cookie Rewriting feature is vulnerable to Lua code injection. An attacker with RBAC permissions to create or modify HTTPProxy resources can craft a malicious value in spec.routes[].cookieRewritePolicies[].pathRewrite.value or spec.routes[].services[].cookieRewritePolicies[].pathRewrite.value that results in arbitrary code execution in the Envoy proxy. The cookie rewriting feature is internally implemented using Envoy's HTTP Lua filter. User-controlled values are interpolated into Lua source code using Go text/template without sufficient sanitization. The injected code only executes when processing traffic on the attacker's own route, which they already control. However, since Envoy runs as shared infrastructure, the injected code can also read Envoy's xDS client credentials from the filesystem or cause denial of service for other tenants sharing the Envoy instance. This vulnerability is fixed in v1.33.4, v1.32.5, and v1.31.6.

LangChain is a framework for building agents and LLM-powered applications. Prior to langchain-text-splitters 1.1.2, HTMLHeaderTextSplitter.split_text_from_url() validated the initial URL using validate_safe_url() but then performed the fetch with requests.get() with redirects enabled (the default). Because redirect targets were not revalidated, a URL pointing to an attacker-controlled server could redirect to internal, localhost, or cloud metadata endpoints, bypassing SSRF protections. The response body is parsed and returned as Document objects to the calling application code. Whether this constitutes a data exfiltration path depends on the application: if it exposes Document contents (or derivatives) back to the requester who supplied the URL, sensitive data from internal endpoints could be leaked. Applications that store or process Documents internally without returning raw content to the requester are not directly exposed to data exfiltration through this issue. This vulnerability is fixed in 1.1.2.

Math.js is an extensive math library for JavaScript and Node.js. From version 13.1.0 to before version 15.2.0, arbitrary JavaScript can be executed via the expression parser of mathjs. This issue has been patched in version 15.2.0.

vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.22.0, an assert-based security check in vLLM's activation function loading allows any unauthenticated attacker to achieve arbitrary code execution on the server by publishing a malicious HuggingFace model, when vLLM runs in Python optimized mode (python -O or PYTHONOPTIMIZE=1). This vulnerability is fixed in 0.22.0.

Improper Input Validation, Improper Control of Generation of Code ('Code Injection') vulnerability in Apache ActiveMQ, Apache ActiveMQ Broker, Apache ActiveMQ All. An authenticated attacker can use the admin web console page to construct a malicious broker name that bypasses name validation to include an xbean binding that can be later used by a VM transport to load a remote Spring XML application. The attacker can then use the DestinationView mbean to send a message to trigger a VM transport creation that will reference this malicious broker name which can lead to loading the malicious Spring XML context file. Because Spring's ResourceXmlApplicationContext instantiates all singleton beans before the BrokerService validates the configuration, arbitrary code execution occurs on the broker's JVM through bean factory methods such as Runtime.exec(). This issue affects Apache ActiveMQ: before 5.19.6, from 6.0.0 before 6.2.5; Apache ActiveMQ Broker: before 5.19.6, from 6.0.0 before 6.2.5; Apache ActiveMQ All: before 5.19.6, from 6.0.0 before 6.2.5. Users are recommended to upgrade to version 6.2.5 or 5.19.6, which fixes the issue.

Rclone is a command-line program to sync files and directories to and from different cloud storage providers. Starting in version 1.48.0 and prior to version 1.73.5, the RC endpoint `operations/fsinfo` is exposed without `AuthRequired: true` and accepts attacker-controlled `fs` input. Because `rc.GetFs(...)` supports inline backend definitions, an unauthenticated attacker can instantiate an attacker-controlled backend on demand. For the WebDAV backend, `bearer_token_command` is executed during backend initialization, making single-request unauthenticated local command execution possible on reachable RC deployments without global HTTP authentication. Version 1.73.5 patches the issue.

protobufjs compiles protobuf definitions into JavaScript (JS) functions. In versions prior to 8.0.1 and 7.5.5, attackers can inject arbitrary code in the "type" fields of protobuf definitions, which will then execute during object decoding using that definition. Versions 8.0.1 and 7.5.5 patch the issue.

Other issue in the WebRTC component. This vulnerability was fixed in Firefox ESR 140.10.2 and Thunderbird 140.10.2.

A flaw was found in ose-openshift-apiserver. This vulnerability allows internal network enumeration, service discovery, limited information disclosure, and potential denial-of-service (DoS) through Server-Side Request Forgery (SSRF) due to missing IP address and network-range validation when processing user-supplied image references.

Insufficient validation of node IDs in Qt SVG module allows arbitrary QML/JavaScript code injection when loading malicious SVG files through the VectorImage component in Qt Quick. While QML execution is typically more restricted than native code execution, this could still lead to denial of service, information disclosure, or other impacts depending on the application's privilege level and data access.

LiteLLM through 2026-04-08 allows remote attackers to execute arbitrary code via bytecode rewriting at the /guardrails/test_custom_code URI.

A command injection vulnerability exists in mlflow/mlflow versions before v3.7.0, specifically in the `mlflow/sagemaker/__init__.py` file at lines 161-167. The vulnerability arises from the direct interpolation of user-supplied container image names into shell commands without proper sanitization, which are then executed using `os.system()`. This allows attackers to execute arbitrary commands by supplying malicious input through the `--container` parameter of the CLI. The issue affects environments where MLflow is used, including development setups, CI/CD pipelines, and cloud deployments.

Improper Input Validation, Improper Control of Generation of Code ('Code Injection') vulnerability in Apache ActiveMQ Broker, Apache ActiveMQ All, Apache ActiveMQ. An authenticated attacker may bypass the fix in CVE-2026-34197 by adding a connector using an HTTP Discovery transport via BrokerView.addNetworkConnector or BrokerView.addConnector through Jolokia if the activemq-http module is on the classpath. A malicious HTTP endpoint can return a VM transport through the HTTP URI which will bypass the validation added in CVE-2026-34197. The attacker can then use the VM transport's brokerConfig parameter to load a remote Spring XML application context using ResourceXmlApplicationContext. Because Spring's ResourceXmlApplicationContext instantiates all singleton beans before the BrokerService validates the configuration, arbitrary code execution occurs on the broker's JVM through bean factory methods such as Runtime.exec(). This issue affects Apache ActiveMQ Broker: before 5.19.6, from 6.0.0 before 6.2.5; Apache ActiveMQ All: before 5.19.6, from 6.0.0 before 6.2.5; Apache ActiveMQ: before 5.19.6, from 6.0.0 before 6.2.5. Users are recommended to upgrade to version 5.19.6 or 6.2.5, which fixes the issue.

A flaw was found in libinput. A local attacker who can place a specially crafted Lua bytecode file in certain system or user configuration directories can bypass security restrictions. This allows the attacker to run unauthorized code with the same permissions as the program using libinput, such as a graphical compositor. This could lead to the attacker monitoring keyboard input and sending that information to an external location.

OpenStack Glance before 29.1.1, 30.x before 30.1.1, and 31.0.0 is affected by Server-Side Request Forgery (SSRF). By use of HTTP redirects, an authenticated user can bypass URL validation checks and redirect to internal services. Only glance image import functionality is affected. In particular, the web-download and glance-download import methods are subject to this vulnerability, as is the optional (not enabled by default) ovf_process image import plugin.

Handlebars provides the power necessary to let users build semantic templates. In versions 4.0.0 through 4.7.8, `Handlebars.compile()` accepts a pre-parsed AST object in addition to a template string. The `value` field of a `NumberLiteral` AST node is emitted directly into the generated JavaScript without quoting or sanitization. An attacker who can supply a crafted AST to `compile()` can therefore inject and execute arbitrary JavaScript, leading to Remote Code Execution on the server. Version 4.7.9 fixes the issue. Some workarounds are available. Validate input type before calling `Handlebars.compile()`; ensure the argument is always a `string`, never a plain object or JSON-deserialized value. Use the Handlebars runtime-only build (`handlebars/runtime`) on the server if templates are pre-compiled at build time; `compile()` will be unavailable.

Happy DOM is a JavaScript implementation of a web browser without its graphical user interface. In versions 15.10.0 through 20.8.7, a code injection vulnerability in `ECMAScriptModuleCompiler` allows an attacker to achieve Remote Code Execution (RCE) by injecting arbitrary JavaScript expressions inside `export { }` declarations in ES module scripts processed by happy-dom. The compiler directly interpolates unsanitized content into generated code as an executable expression, and the quote filter does not strip backticks, allowing template literal-based payloads to bypass sanitization. Version 20.8.8 fixes the issue.

A flaw was found in Red Hat Quay's Proxy Cache configuration feature. When an organization administrator configures an upstream registry for proxy caching, Quay makes a network connection to the specified registry hostname without verifying that it points to a legitimate external service. An attacker with organization administrator privileges could supply a crafted hostname to force the Quay server to make requests to internal network services, cloud infrastructure endpoints, or other resources that should not be accessible from the Quay application.

Handlebars provides the power necessary to let users build semantic templates. In versions 4.0.0 through 4.7.8, the `@partial-block` special variable is stored in the template data context and is reachable and mutable from within a template via helpers that accept arbitrary objects. When a helper overwrites `@partial-block` with a crafted Handlebars AST, a subsequent invocation of `{{> @partial-block}}` compiles and executes that AST, enabling arbitrary JavaScript execution on the server. Version 4.7.9 fixes the issue. Some workarounds are available. First, use the runtime-only build (`require('handlebars/runtime')`). The `compile()` method is absent, eliminating the vulnerable fallback path. Second, audit registered helpers for any that write arbitrary values to context objects. Helpers should treat context data as read-only. Third, avoid registering helpers from third-party packages (such as `handlebars-helpers`) in contexts where templates or context data can be influenced by untrusted input.

Handlebars provides the power necessary to let users build semantic templates. In versions 4.0.0 through 4.7.8, the Handlebars CLI precompiler (`bin/handlebars` / `lib/precompiler.js`) concatenates user-controlled strings — template file names and several CLI options — directly into the JavaScript it emits, without any escaping or sanitization. An attacker who can influence template filenames or CLI arguments can inject arbitrary JavaScript that executes when the generated bundle is loaded in Node.js or a browser. Version 4.7.9 fixes the issue. Some workarounds are available. First, validate all CLI inputs before invoking the precompiler. Reject filenames and option values that contain characters with JavaScript string-escaping significance (`"`, `'`, `;`, etc.). Second, use a fixed, trusted namespace string passed via a configuration file rather than command-line arguments in automated pipelines. Third, run the precompiler in a sandboxed environment (container with no write access to sensitive paths) to limit the impact of successful exploitation. Fourth, audit template filenames in any repository or package that is consumed by an automated build pipeline.