CVE & CISA-KEV Catalog

LLaMA-Factory through 0.9.5 contains a remote code execution vulnerability that allows attackers with WebUI access to execute arbitrary Python code by supplying a malicious model path in the Chat or Training interfaces. The application passes user-supplied model path input unvalidated into AutoTokenizer.from_pretrained() and AutoModel.from_pretrained() with a hardcoded trust_remote_code=True parameter, causing the Hugging Face transformers library to fetch and execute arbitrary code from a remote or local model repository with the privileges of the server process.

Published 2026-06-30

Improper handling of untrusted remote references in Snowflake CLI versions prior to 3.19 allowed server-side request forgery. The SQL statement reader's !source/!load directives could reference remote URLs that were retrieved at runtime without sufficient restriction on the request destination. By supplying crafted SQL content processed through a vulnerable command path, an attacker could cause the victim's environment to issue unintended outbound requests to internal or otherwise non-public network locations, and could cause remote SQL content to be retrieved and executed in the context of the victim user's session. Successful exploitation requires the victim to process attacker-controlled content through a vulnerable command path and is limited by the privileges available to that session

Published 2026-06-29

pnpm is a package manager. Prior to 10.34.2 and 11.5.3, pnpm can persist package-manager bootstrap metadata in the first YAML document of pnpm-lock.yaml. Before the patch, direct pnpm execution trusted an already resolved packageManagerDependencies entry when the committed env lockfile contained matching pnpm and @pnpm/exe versions. A malicious repository could therefore commit package-manager lockfile package records and snapshots that bypassed fresh package-manager resolution, then cause pnpm to install and execute bytes selected by that committed lockfile state during automatic version switching. This vulnerability is fixed in 10.34.2 and 11.5.3.

Published 2026-06-25

pnpm is a package manager. Prior to 10.34.2 and 11.5.3, pnpm can install configDependencies declared in pnpm-workspace.yaml before command dispatch. Before the patch, a repository could declare pacquet or @pnpm/pacquet as a config dependency and pnpm treated that repository-controlled dependency as an install-engine opt-in. During install, pnpm resolved a platform-specific @pacquet/<platform>-<arch>/pacquet binary from node_modules/.pnpm-config/<packageName> and spawned it as the developer or CI user. This vulnerability is fixed in 10.34.2 and 11.5.3.

Published 2026-06-25

pnpm is a package manager. Prior to 10.34.2 and 11.5.3, the generic peer-suffix normalizer also stripped parenthesized text from git, URL, tarball, file, and other opaque locators. Approval for one source string could therefore authorize a different attacker-controlled source whose locator normalized to the same value. This vulnerability is fixed in 10.34.2 and 11.5.3.

Published 2026-06-25

Pi is a minimal terminal coding harness. Pi before 0.79.0 loaded project-local configuration and resources from a repository's .pi directory without first asking the user to trust that repository. This included project-local extensions, which are executable TypeScript or JavaScript modules loaded into the Pi process. An attacker who controls a repository could place Pi-specific project resources in that repository. If a user then started Pi from that working tree, the project-local extension code could run with the same privileges as the local Pi process without the user having a convenient way to make a trust decision. This vulnerability is fixed in 0.79.0.

Published 2026-06-23

MISP allowed an authenticated site administrator to set the Kafka_rdkafka_config setting to an arbitrary filesystem path. MISP subsequently parsed the referenced INI file and passed its options to rdkafka. A crafted attacker-controlled configuration file could use rdkafka options such as plugin.library.paths to load an external library, resulting in arbitrary code execution with the privileges of the MISP process. An attacker could leverage a MISP-writable location, such as an uploaded file or administrative image, to host the malicious configuration file. The issue is fixed by restricting the setting to absolute .ini files located only in approved configuration directories outside the webroot and MISP upload targets.

Published 2026-06-22

In Eclipse Theia versions prior to 1.71.0, files matching the pattern .prompts/*.prompttemplate in a workspace were automatically loaded and could override or extend the AI agent's system prompts. An attacker could craft a malicious repository containing prompt template files that, when the workspace was opened in Theia, replaced the AI's system instructions with attacker-controlled content (indirect prompt injection). Combined with other AI chat features available in untrusted workspaces, this enabled attack chains leading to data exfiltration via Markdown image rendering or arbitrary command execution via task definitions.

Published 2026-06-18

In Eclipse Theia versions prior to 1.69.0, custom task definitions in workspace files (e.g. .theia/tasks.json, .vscode/tasks.json) could be executed without requiring workspace trust. An attacker could craft a malicious repository that, when cloned and opened in Theia, leads to execution of arbitrary commands with the user's privileges. In combination with AI chat features and a workspace .theia/settings.json that disabled tool confirmation, this could be triggered automatically by sending a message in the AI chat.

Published 2026-06-18

In Eclipse Theia versions prior to 1.71.0, the AI chat agent processed workspace file and directory names as part of its prompt context without distinguishing them from system instructions. An attacker could craft a malicious repository with adversarial directory or file names that, when analyzed by the AI agent, would cause the agent to follow attacker-controlled instructions (indirect prompt injection). Combined with other AI chat features available in untrusted workspaces, this enabled attack chains leading to data exfiltration via Markdown image rendering or arbitrary command execution via task definitions.

Published 2026-06-18

In Eclipse Theia versions prior to 1.71.0, the AI chat rendered Markdown image tags from AI responses, triggering HTTP requests to arbitrary external URLs without restriction. Combined with prompt injection in a malicious workspace, an attacker could induce the AI agent to construct image URLs encoding sensitive information from the workspace or conversation context, exfiltrating it to attacker-controlled servers. The workspace trust enforcement introduced in v1.71.0 mitigates the documented attack chain by disabling AI features in untrusted workspaces.

Published 2026-06-18

Dell PowerFlex Manager, version(s) prior to 5.1.0.1, contain(s) an Inclusion of Functionality from Untrusted Control Sphere vulnerability. An unauthenticated attacker with remote access could potentially exploit this vulnerability, leading to Information disclosure.

Published 2026-06-17

Yeoman Environment provides an API to discover, create, and run generators, and to configure where and how a generator is resolved. Versions 2.9.0 through 6.0.0 install missing local generator packages from caller-supplied package names without user confirmation. In downstream consumers that pass attacker-controlled project configuration into this path, this can result in arbitrary package installation and code execution during CLI bootstrap. The vulnerable method is installLocalGenerators(), which calls repository.install() directly without prompting the user. This issue has been fixed in version 6.0.0.

Published 2026-06-16

Cursor is a code editor built for programming with AI. In versions prior to 3.0.0, the Cursor Desktop could execute workspace-defined Claude hook commands from .claude/settings.local.json without dedicated user approval. A malicious workspace or agent-created file could configure hooks that run local commands in the user's context when an agent turn ends. This could allow sandbox escape, persistence across turns, local data access, or follow-on compromise. This issue has been fixed in version 3.0.0.

Published 2026-06-15

When the application executes the JavaScript script embedded in the PDF within the sandbox, it fails to intercept some dangerous interfaces, which allows remote scripts to be loaded, resulting in arbitrary code execution.

Published 2026-06-15

Vim is an open source, command line text editor. Prior to version 9.2.0561, the Python omni-completion script in python3complete.vim for Vim with the +python3 interpreter enabled (and the legacy pythoncomplete.vim for builds with the +python interpreter) executes the import and from statements found in the current buffer through Python's import machinery. Because the buffer's working directory is on sys.path, opening a hostile .py file with a sibling Python package and invoking omni-completion runs that package's top-level code as the editing user. This issue has been patched in version 9.2.0561.

Published 2026-06-13

OpenClaw before 2026.5.18 contains a code execution vulnerability where marketplace runtime extension metadata can redirect loading toward unscanned package payloads. Attackers with trusted operator access can manipulate extension metadata to load plugin code outside reviewed package entry points, bypassing security scanning.

Published 2026-06-11

In Duck Site before version 1.0.1, the repository has a deploy workflow that runs after the build workflow completes. The build workflow runs on pull requests, while the deploy workflow runs with package-write permissions and deployment secrets. If an attacker can make a pull request build satisfy the deploy workflow’s main branch condition, the deploy job checks out the triggering workflow commit, builds it into a Docker image, pushes it as latest, and triggers Dokploy deployment. This can allow attacker-controlled pull request code to become the deployed production site image without being merged. This issue has been patched in version 1.0.1.

Published 2026-06-11

Quest Bot is an opensource modern Discord Bot built for moderation, utilities and support. Prior to version 1.0.3, the repository has a privileged deploy workflow that runs after the unprivileged build workflow completes. The build workflow runs on pull requests, and the deploy workflow checks out the triggering workflow’s head_sha, builds that code into a Docker image, pushes it as latest, and triggers production deployment. If an attacker can open a pull request from a branch named main, the deploy workflow condition can treat the PR build as deployable and build the attacker-controlled commit in a privileged deployment context. This can result in malicious container deployment and production bot compromise. This issue has been patched in version 1.0.3.

Published 2026-06-11

Atril Document Viewer is the default document reader of the MATE desktop environment for Linux. A single-click remote code execution vulnerability in versions prior to 1.26.3 and 1.28.4 allows an attacker to achieve arbitrary code execution as the user by tricking them into clicking a link inside a malicious PDF document. The PDF can be packaged as a polyglot file that is simultaneously a valid PDF and a valid ELF shared library, making the attack a single-file, single-click, configuration-independent RCE on stock atril installations. The root cause is `shell/ev-application.c:ev_spawn`, which builds a command line from attacker-controlled PDF link-destination fields without applying `g_shell_quote`. The cmdline is then handed to `g_app_info_create_from_commandline`, which shell-parses it b

Published 2026-06-10

Inclusion of functionality from untrusted control sphere in Visual Studio Code allows an unauthorized attacker to elevate privileges locally.

Published 2026-06-09

This CVE was assigned by Chrome. Microsoft Edge (Chromium-based) ingests Chromium, which addresses this vulnerability. Please see Google Chrome Releases for more information.

Published 2026-06-05

Version 3.0.7 of the Securly Chrome Extension dynamically registers content13.min.js as a content script via chrome.scripting.registerContentScripts() at runtime. This script is NOT declared in manifest.json and bypasses Chrome Web Store static security review. It runs on all URLs and immediately hides all page content, creates a full-page overlay, pauses all videos, and only restores content when the service worker confirms the page passes filtering. If Securly's servers are unreachable, pages remain indefinitely hidden.

Published 2026-06-03

A vulnerability in the LightGlue model loading path of huggingface/transformers version 5.2.0 allows an attacker-controlled model repository to execute arbitrary code during model initialization. The issue arises because the `trust_remote_code` parameter, intended to prevent remote code execution, is overridden by untrusted serialized configuration data in a nested code path. Specifically, when loading a LightGlue model using `AutoModel.from_pretrained()` with `trust_remote_code=False`, the `LightGlueConfig` reads the `trust_remote_code` value from the untrusted `config.json` file and propagates it into nested `AutoConfig.from_pretrained()` calls. This results in the execution of attacker-provided Python modules, even when the victim explicitly disables remote code execution. The vulnerabi

Published 2026-06-03

An inclusion of functionality from untrusted control sphere vulnerability in MinGW DLL component in Synology Hyper Backup Explorer before 3.0.1-0156 allows local users to execute arbitrary code via unspecified vectors.

Published 2026-06-03

An inclusion of functionality from untrusted control sphere vulnerability in OpenSSL configuration in Synology Active Backup for Business Recovery Media Creator before 2.5.0-2081 allows local users to execute arbitrary code via unspecified vectors.

Published 2026-06-03

Espressif Shared GitHub DangerJS is a reusable GitHub Action CI DangerJS workflow for Espressif GitHub projects. Prior to 1.0.1, the action's entrypoint.sh invoked DangerJS from the caller's workspace after copying the fork's checkout into it, creating an untrusted search path for both binary resolution and Node.js module resolution. A fork pull request processed by a pull_request_target workflow could therefore cause fork-supplied code to execute inside the action container in place of the action's own code. This vulnerability is fixed in 1.0.1.

Published 2026-05-28

The MLX inference backend in Docker Model Runner on macOS uses the MLX-LM library, which unconditionally imports and executes arbitrary Python files from model directories via the model_file configuration field in config.json. When a model's config.json specifies a model_file pointing to a Python file, MLX-LM uses importlib to load and execute it with no trust_remote_code gate or equivalent safety check. The MLX backend runs without sandboxing, resulting in arbitrary code execution on the Docker host as the Docker Desktop user. Any container on the Docker network can trigger this by calling the model-runner.docker.internal API to pull a malicious model from an attacker-controlled OCI registry and request inference.

Published 2026-05-22

The vllm-metal inference backend in Docker Model Runner on macOS unconditionally sets trust_remote_code=True when loading model tokenizers, and runs without sandboxing. This causes transformers.AutoTokenizer.from_pretrained() to import and execute arbitrary Python files included in any model pulled from an OCI registry, resulting in arbitrary code execution on the Docker host as the Docker Desktop user when inference is triggered. Any container on the Docker network can trigger this by calling the model-runner.docker.internal API to pull a malicious model and request inference.

Published 2026-05-22

Concrete CMS 9.5.0 and below emits a CSRF token in the local_available_update.php view ($token->output('do_update')) but the corresponding do_update() method in concrete/controllers/single_page/dashboard/system/update/update.php never calls $this->token->validate('do_update'). The form is rendered as a POST form, meaning the token reaches the browser, but because the controller discards it without verification, an attacker can craft a cross-site POST that triggers a core CMS update to an attacker-specified version string. In order to be vulnerable, theictim must be passing canUpgrade()anda valid update version must be present under DIR_CORE_UPDATES. The Concrete CMS security team gave this vulnerability a CVSS v.4.0 score of 7.5 with vector CVSS:4.0/AV:N/AC:H/AT:P/PR:N/UI:A/VC:H/VI:H/VA:H

Published 2026-05-21

Concrete CMS 9.5.0 and below does not validate a CSRF token before processing requests to /dashboard/extend/update/prepare_remote_upgrade/<remoteMPID>. An attacker who controls the remote package returned for a known marketplace item ID can overwrite the package PHP on disk and force its upgrade() method to execute in a single browser navigation. This results in remote code execution as the web server user. In order to be vulnerable, the victim must be passing canInstallPackages, victim site must be connected to the Concrete marketplace; and the attacker controls the package returned for a marketplace item ID already installed on the victim site. The Concrete CMS security team gave this vulnerability a CVSS v.4.0 score of 7.5 with vector CVSS:4.0/AV:N/AC:H/AT:P/PR:N/UI:A/VC:H/VI:H/VA:H/S

Published 2026-05-21

pip prior to version 26.1 would run self-update check functionality after installing wheel files which required importing well-known Python modules names. These module imports were intentionally deferred to increase startup time of the pip CLI. The patch changes self-update functionality to run before wheels are installed to prevent newly-installed modules from being imported shortly after the installation of a wheel package. Users should still review package contents prior to installation.

Published 2026-05-19

Rapid7 Metasploit Pro is vulnerable to a local privilege escalation attack that allows a user to gain SYSTEM level control of a Windows host. When started the metasploitPostgreSQL service would start the postgres.exe child process which would in turn load an OpenSSL configuration file from a static location. This static location would be writable by a pre-existing "vagrant" user, if they already existed on the system. Metasploit does not create local accounts, an Administrator would need to create it. By planting a crafted openssl.cnf file an attacker can trick the high-privilege service into executing arbitrary commands. This effectively permits the unprivileged vagrant user to bypass security controls and achieve a full host compromise under the agent's SYSTEM level access.

Published 2026-05-15

css_parser is a Ruby CSS parser. Prior to 2.1.0 and 1.22.0, the CSS Parser gem does not validate HTTPS connections, allowing a Man-in-the-Middle (MITM) attacker to inject or modify CSS content when stylesheets are loaded via HTTPS. The connection is established with OpenSSL::SSL::VERIFY_NONE, meaning any HTTPS certificate—even entirely untrusted—will be accepted without validation. This vulnerability is fixed in 2.1.0 and 1.22.0.

Published 2026-05-14

PyTorch Lightning is a deep learning framework to pretrain and finetune AI models. Versions 2.6.2 and 2.6.2 have introduced functionality consistent with a credential harvesting mechanism.

Published 2026-05-14

vm2 is an open source vm/sandbox for Node.js. Prior to 3.11.0, NodeVM's builtin allowlist can be bypassed when the module builtin is allowed (including via the '*' wildcard). The module builtin exposes Node's Module._load(), which loads any module by name directly in the host context, completely bypassing vm2's builtin restriction. This allows sandboxed code to load excluded builtins like child_process and achieve remote code execution. This vulnerability is fixed in 3.11.0.

Published 2026-05-13

OpenClaw before 2026.4.20 contains an improper environment variable validation vulnerability in MCP stdio server configuration that allows attackers to execute arbitrary code. Malicious workspace configurations can pass dangerous startup variables like NODE_OPTIONS, LD_PRELOAD, or BASH_ENV to spawned MCP server processes, enabling code injection when operators start sessions using those servers.

Published 2026-05-11

Kdenlive before 26.04.1 allows dangerous proxy parameters when an attacker-controlled project file is used.

Published 2026-05-09

PraisonAI is a multi-agent teams system. Prior to version 4.6.34, PraisonAI's MCP (Model Context Protocol) server (praisonai mcp serve) registers four file-handling tools by default — praisonai.rules.create, praisonai.rules.show, praisonai.rules.delete, and praisonai.workflow.show. Each accepts a path or filename string from MCP tools/call arguments and joins it onto ~/.praison/rules/ (or, for workflow.show, accepts an absolute path) with no containment check. The JSON-RPC dispatcher passes params["arguments"] blind to each handler via **kwargs without validating against the advertised input schema. By setting rule_name="../../<some-path>" an attacker walks out of the rules directory and writes any file the running user can write. Dropping a Python .pth file into the user site-packages dir

Published 2026-05-08

electerm is an open-sourced terminal/ssh/sftp/telnet/serialport/RDP/VNC/Spice/ftp client. From versions 3.0.6 to before 3.8.15, electerm is vulnerable to arbitrary local code execution via deep links, CLI --opts, or crafted shortcuts. Exploit requires clicking a crafted electerm://... link or opening a crafted shortcut/command that launches electerm with attacker-controlled opts. This issue has been patched in version 3.8.15.

Published 2026-05-08

electerm is an open-sourced terminal/ssh/sftp/telnet/serialport/RDP/VNC/Spice/ftp client. Prior to version 3.7.16, the runWidget function in src/app/widgets/load-widget.js constructs a file path by directly concatenating user‑supplied widget identifiers without any sanitisation. Because runWidget is exposed to the renderer process via an asynchronous IPC handler with no input validation, an attacker who achieves JavaScript execution inside the renderer (for example, through a malicious plugin or a cross‑site scripting flaw in the built‑in webview) can abuse a path traversal (../) to load and execute an arbitrary JavaScript file anywhere on the victim’s filesystem. This gives the attacker local code execution with the full privileges of the electerm process, leading to complete system compr

Published 2026-05-08

OpenClaw before 2026.4.10 contains a plugin trust bypass vulnerability that allows channel setup catalog lookups to resolve workspace plugin shadows before bundled channel plugins. Attackers can exploit this by crafting malicious workspace plugins that bypass intended trust gates during setup-time plugin loading.

Published 2026-05-05

OpenClaw before 2026.4.9 contains an authentication bypass vulnerability allowing untrusted workspace plugins to be auto-enabled during non-interactive onboarding when provider auth choices are shadowed. Attackers can exploit this by crafting malicious workspace plugins that are automatically selected and enabled during authentication setup without explicit user consent.

Published 2026-05-05

An issue was discovered in OpenStack ironic-python-agent 1.0.0 through 11.5.0. Ironic Python Agent (IPA) sometimes executes grub-install from within a chroot of the deployed partition image, leading to code execution in the case of a malicious image.

Published 2026-05-01

OpenClaw before 2026.3.31 allows workspace .env files to override the OPENCLAW_BUNDLED_PLUGINS_DIR environment variable, compromising plugin trust verification. Attackers with control over workspace configuration can inject malicious plugins by overriding the bundled plugin trust root directory.

Published 2026-04-28

OpenStack Ironic before 35.0.1 allows ipmitool execution in a non-default configuration that has a console interface.

Published 2026-04-28

OpenClaw before 2026.3.28 contains an arbitrary code execution vulnerability in mirror mode that converts untrusted sandbox files into workspace hooks. Attackers with mirror mode access can execute arbitrary code on the host during gateway startup by exploiting enabled workspace hooks.

Published 2026-04-23

OpenClaw before 2026.3.31 allows workspace .env files to override the OPENCLAW_BUNDLED_HOOKS_DIR environment variable, enabling loading of attacker-controlled hook code. Attackers can replace trusted default-on bundled hooks from untrusted workspaces to execute arbitrary code.

Published 2026-04-23

A flaw was found in InstructLab. The `linux_train.py` script hardcodes `trust_remote_code=True` when loading models from HuggingFace. This allows a remote attacker to achieve arbitrary Python code execution by convincing a user to run `ilab train/download/generate` with a specially crafted malicious model from the HuggingFace Hub. This vulnerability can lead to complete system compromise.

Published 2026-04-22

goshs is a SimpleHTTPServer written in Go. Prior to 2.0.0-beta.6, goshs has an ArtiPACKED vulnerability. ArtiPACKED can lead to leakage of the GITHUB_TOKEN through workflow artifacts, even though the token is not present in the repository source code. This vulnerability is fixed in 2.0.0-beta.6.

Published 2026-04-21