Browse / AI Framework and Platform Flaws
AI Framework and Platform Flaws
Classic software bugs in the libraries and platforms that run AI: memory corruption, traversal, SSRF, auth. 1,542 records. Reference: CWE.
- Denial of service 377
- Authentication and access control 256
- Path traversal and file access 192
- Memory corruption 159
- Code and command execution 133
- Server-side request forgery 103
- Information exposure 78
- Sandbox escape 76
- Cross-site scripting 68
- Other weaknesses 46
- SQL and query injection 22
- Real-world incidents 10
- Cross-site request forgery 8
- Unsafe deserialisation 8
- Data exfiltration 6
- PaddlePaddle floating point exception in paddle.linalg.eig A floating-point exception in the eigendecomposition routine crashes the process when given crafted input.
- PaddlePaddle stack overflow in paddle.linalg.lu_unpack Crafted input to the linear algebra unpacking function overflows the stack in the PaddlePaddle framework.
- PaddlePaddle command injection in get_online_pass_interval Unsanitised input to get_online_pass_interval is interpolated into an operating system command, allowing arbitrary commands to run.
- PaddlePaddle floating point exception in paddle.linalg.matrix_rank A floating point exception in the matrix rank routine crashes the process when given crafted input.
- PaddlePaddle segfault in paddle.mode An out-of-bounds read inside paddle.mode crashes the process when crafted input is supplied.
- PaddlePaddle nullptr dereference in paddle.crop A null pointer dereference in the crop operation crashes PaddlePaddle, letting anyone who can call it halt the process.
- PaddlePaddle command injection in _wget_download The _wget_download helper builds a shell command from unsanitised input, letting an attacker run arbitrary operating system commands.
- PaddlePaddle floating point exception in paddle.lerp Crafted inputs to paddle.lerp trigger a floating point exception that crashes the process.
- PaddlePaddle floating point exception in paddle.topk Calling paddle.topk with certain inputs triggers a floating point exception that crashes the process.
- PaddlePaddle floating point exception in paddle.amin A crafted input to paddle.amin triggers a floating point exception that crashes the process.
- PaddlePaddle segfault in paddle.dot A null pointer in paddle.dot causes the process to crash when given crafted input.
- PaddlePaddle floating point exception in paddle.nanmedian Specific inputs to paddle.nanmedian trigger a floating point exception that crashes the process.
- Gradio makes the `/file` secure against file traversal and server-side request forgery attacks Raw HTTP tools could read arbitrary server files and make server-side requests through Gradio's file route, which browser-style URLs could not reach.
- MLflow Server-Side Request Forgery (SSRF) Crafted requests make the MLflow server fetch attacker-chosen internal HTTP endpoints, which could lead to code execution on cloud hosts.
- MLflow Path Traversal Vulnerability Filenames containing backslash-dot traversal sequences let a request escape the intended directory and reach other files on the server.
- MLFlow Path Traversal Vulnerability A path traversal flaw in MLflow lets a user reach files outside the intended directory, leading to command execution on the host.
- MLflow Local File Disclosure Vulnerability The MLflow server lets users read arbitrary local files on the host, exposing sensitive server data.
- MLflow Path Traversal Vulnerability The server accepts uploads that write arbitrary files to arbitrary locations on the remote filesystem, using the server process's permissions.
- GitHub Security Lab (GHSL) Vulnerability Report: Arbitary write GHSL-2023-182 A file write operation accepts unvalidated paths, letting attackers create or overwrite files outside the intended directory.
- Path traversal in MLflow A filename containing backslash-dot sequences escapes the intended directory during file handling in MLflow.
- Gradio Exposure of Sensitive Information to an Unauthorized Actor vulnerability Gradio exposes sensitive information to actors who should not receive it; the advisory gives no further technical detail.
- External Control of File Name or Path in h2oai/h2o-3 Unauthenticated requests can make the h2o export write to attacker-chosen paths, overwriting arbitrary server files with attacker-shaped content.
- Path traversal in MLflow Crafted path arguments in MLflow escape the intended directory, letting requests read or write files outside it.
- Server-Side Request Forgery in mindsdb The file upload endpoint fetches a user-supplied URL without validation, letting requests reach internal hosts.
- Improper Input Validation in mindsdb A user-supplied name is used unvalidated in a temporary file path, letting a request write files to arbitrary server locations.
- Jinja2 template injection in mlflow User-controlled content reaches a Jinja2 template without neutralisation, so template expressions can be evaluated and run code on the server.
- Cross-site Scripting (XSS) in MLflow The Content-Type request header is echoed into an HTML response without escaping, so a crafted header runs script in the victim's browser.
- Information exposure in MLflow A crafted request to the MLflow REST API lets a remote attacker retrieve sensitive information from versions up to 2.8.1.
- Ray has arbitrary code execution via jobs submission API The Ray job submission API lets a remote party run arbitrary code on the cluster, which the vendor says is only meant for controlled networks.
- TorchServe ZipSlip Model archives uploaded through the management API are extracted without checking entry paths, so crafted zip entries can write outside the target directory.
- H2O local file inclusion vulnerability Unauthenticated requests to the import and parse setup endpoints let callers read arbitrary files the server process can access.
- MLflow authentication requirement bypass can allow a user to arbitrarily create an account A flaw lets an unauthenticated caller create accounts despite the server's requirement that registration be authenticated.
- Remote Code Execution due to Full Controled File Write in mlflow Unrestricted file writes on the MLflow web server let an attacker overwrite files such as shell startup scripts to run commands.
- MLflow allowed arbitrary files to be PUT onto the server MLflow accepts file uploads written to arbitrary server paths without restricting the destination to an allowed directory.
- Ray Path Traversal vulnerability The unauthenticated log API endpoint accepts file paths without restriction, letting callers read arbitrary files on the server.
- Ray Missing Authorization vulnerability The static file route accepts traversal sequences, letting unauthenticated requests read arbitrary files on the server.
- Ray OS Command Injection vulnerability A URL parameter on the Ray dashboard's cpu_profile endpoint is passed into an OS command, letting unauthenticated remote users run commands.
- Label Studio Object Relational Mapper Leak Vulnerability in Filtering Task Filter paths that traverse Django relations let the task filter reveal sensitive fields such as password hashes one character at a time.
- Label Studio has Hardcoded Django `SECRET_KEY` that can be Abused to Forge Session Tokens A hardcoded Django secret key lets anyone sign valid session tokens, impersonating accounts once password hashes are leaked.
- LangChain Server Side Request Forgery vulnerability Crawling in LangChain's recursive URL document loader can be pointed from an external page to internal servers, so the host issues requests it should not.
- TorchServe Pre-Auth Remote Code Execution An outdated SnakeYAML library lets crafted YAML instantiate Java objects in TorchServe, enabling code execution without authentication.
- TorchServe Server-Side Request Forgery vulnerability The model loading API fetches any supplied URL over HTTP and writes it to disk when the default allowed-URL setting is in place.
- Gradio arbitrary file upload vulnerability Gradio's /upload route accepts arbitrary files without restricting type or destination.
- Weaviate denial of service vulnerability A type conversion error in the vector database server causes denial of service in releases up to 1.20.0.
- MindsDB can be made to not verify SSL certificates MindsDB code paths disable certificate checks, so traffic to connected services can be intercepted without detection.
- mlflow vulnerable to OS Command Injection Unsanitised input reaches an operating system command in MLflow builds before 2.6.0, letting an attacker run shell commands.
- Use after free in PaddlePaddle A buffer is used after being freed in the paddle.diagonal tensor operation, a use-after-free in native code.
- Command injection in PaddlePaddle Unsanitised input reaches an operating system command in a filesystem helper, letting an attacker run arbitrary commands.
- Float point exception (FPE) in paddlepaddle A floating point exception in matrix_power on certain inputs crashes the process.
- Heap buffer overflow in PaddlePaddle A heap buffer overflow in the paddle.trace tensor operation lets crafted input corrupt memory, causing crashes or possibly further compromise.