Mercredi 23 septembre 2026

OpenC3 cosmos

OpenC3 COSMOS provides the functionality needed to send commands to and receive data from one or more embedded systems. From 5.12.0 until 7.3.0, an authenticated actor can write the pypi_url setting through set_setting at POST /openc3-api/api, then cause OpenC3::PluginModel.install_phase2 in openc3/lib/openc3/models/plugin_model.rb to interpolate the value into a shell command while installing a plugin with Python dependency metadata. Shell metacharacters in the setting are interpreted by the command shell, allowing arbitrary operating-system commands to run as the openc3 service user with access to Redis and bucket credentials. Open-source deployments permit any authenticated user to reach the affected operations, while Enterprise deployments require an administrator. This issue is fixed in version 7.3.0.

Pollen Robotics Reachy Mini

The Reachy Mini daemon exposes an HTTP API for managing the robot. Its app installation endpoint, POST /apps/install in src/reachy_mini/daemon/app/routers/apps.py, has no authentication. The handler's only dependency is Depends(get_app_manager), which just hands back the manager object from application state, so nothing in the chain ever checks a credential. The endpoint takes an AppInfo body naming a Hugging Face Space. The daemon downloads that Space and installs it as a Python package through install_package in src/reachy_mini/apps/sources/local_common_venv.py, using uv or pip. Installing a Python package runs the package's own build and setup code, so whoever chooses the Space chooses what code the robot runs. Anyone can publish a public Hugging Face Space, so this is not a meaningful restriction on the attacker. How far this reaches depends on the model. In _resolve_bind_host in src/reachy_mini/daemon/app/main.py the daemon binds 0.0.0.0 when it runs as the wireless version and 127.0.0.1 otherwise, with the vendor's own comment explaining that the robot has to be reachable on the LAN. On a wireless unit, then, any host on the same network can install and run code on the robot without credentials. One related change has already shipped but does not fix this. Version 1.8.2 replaced the wildcard CORS policy with an allow list of localhost and Tauri origins. That closes the browser drive-by route, where a web page the victim visits silently calls the endpoint in the background. It has no effect on this issue: CORS is enforced by browsers and governs whether script may read a response, while a direct HTTP request from another machine on the network involves no browser, no preflight and no CORS check at all.

Hugging Face Transformers

A vulnerability in Hugging Face Transformers (versions 4.57.0 to 5.16.1) allows remote Python files to be written to local disk without user consent when using GenerativePreTrainedModel.load_custom_generate(). The function fetches and caches a remote module file before performing the required trust_remote_code consent check, inverting the security model enforced by other code-loading paths (such as AutoConfig, AutoModel, and AutoTokenizer). As a result, attacker‑controlled Python code from custom_generate/generate.py is copied into the user’s ~/.cache/huggingface/modules directory even if the user declines the trust prompt. Although execution is correctly gated, the file write is not reversible and can persist across sessions. This can lead to persistent, unauthorized files on disk and stale cache collisions where cached attacker code may later be executed during trusted model loads. The issue stems from an unconditional file write in dynamic_module_utils.py prior to any trust verification.

psd-tools

psd-tools is a Python package for working with Adobe Photoshop PSD files. Prior to 1.17.4, PSDImage.composite() and PSDImage.numpy() allocated output buffers from attacker-controlled PSD header geometry, including width, height, channels, depth, and per-layer rectangles, before validating those values against the available file data. A tiny crafted PSD could therefore cause multi-gigabyte memory allocation, and PSDImage.composite() could return a black image with only a warning instead of raising an exception. Services that composite untrusted PSD files could be terminated by out-of-memory handling. This issue is fixed in version 1.17.4.

dataease SQLBot

SQLBot is an intelligent Text-to-SQL system based on large language models and RAG. Prior to 1.9.0, the POST /api/v1/datasource/parseExcel endpoint in backend/apps/datasource/api/datasource.py uses attacker-controlled multipart filename data when selecting where uploaded content is stored, writes the content before spreadsheet parsing and validation finish, and can transform a double-extension filename into a Python source file. An attacker able to submit a crafted multipart upload can use these behaviors to place attacker-controlled content in /opt/sqlbot/app/alembic/versions/ even when a spreadsheet parsing failure after the file write causes the endpoint to return an error. The planted file remains on disk, and subsequent SQLBot startup or migration processing causes Alembic to import the module and execute its module-level statements in the SQLBot application runtime. This issue is fixed in version 1.9.0.

axolotl-ai-cloud axolotl

Axolotl before 0.19.0 contains a remote code execution vulnerability in the multipack patch path where trust_remote_code defaults to None instead of False, causing the security guard to be bypassed. Attackers can execute arbitrary Python code by crafting a malicious Hugging Face model repository selected as base_model, which is loaded with hardcoded trust_remote_code=True during AutoModelForCausalLM.from_pretrained.

dataelement bisheng

BISHENG before 2.6.0 contains a remote code execution vulnerability in the workflow run_once endpoint that allows authenticated users to execute arbitrary Python code. Attackers can submit crafted Code node definitions to the POST /api/v1/workflow/run_once endpoint, which executes them with exec() without sandboxing, gaining access to filesystem, credentials, and internal network resources.

AsyncFuncAI deepwiki-open

DeepWiki-Open through commit d92819a contains an arbitrary file read vulnerability in the unauthenticated /ws/chat WebSocket endpoint that accepts repo_url as a filesystem path with no containment. Attackers can supply arbitrary directory paths to read all files with supported extensions including Python, JavaScript, YAML, and JSON files containing hardcoded secrets and credentials.

Crypt::SelfCertificate versions from 1.01 through 1.05 for Perl contains malware which executes Python code from an obfuscated URL. The generate_certificate runs a Python script saved as a certificate file. The pyhton script attempts to retrieve code from a hardcoded http URL that is obfuscated with base64 encoding and run the response body directly. The impact is that arbitrary code can be invoked as the user, without a dropped script being saved on the affected host. The releases have no test scripts nor build hooks. The intention may have been to trigger the payload after installation. For version 1.01, the dropper script is in lib/Crypt/SelfCertificate/sample/validate.p12. For version 1.05, the dropper script is in lib/Crypt/SelfCertificate/sample/cert7.pem. The SHA-256 digests of the files are fbff21f45ff748365062a5e36fb2d72558cad82a507a6f357f320b4fcdf07760 Crypt-SelfCertificate-1.01.tar.gz 27b2d2d3174ad771474fff2521f5084ec231e9218ea8c832515aef1cbd5897bc lib/Crypt/SelfCertificate/sample/validate.p12 9fdfa7d69b034b77d4510cda567e8da1e486ca81c7daaadc5732a45c41d71991 Crypt-SelfCertificate-1.05.tar.gz 27b2d2d3174ad771474fff2521f5084ec231e9218ea8c832515aef1cbd5897bc lib/Crypt/SelfCertificate/sample/cert7.pem