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
- MLflow allows low privilege users to delete any artifact Users with only edit permission on an experiment can send DELETE requests that remove artifacts, which the documented permissions do not allow.
- MLflow has a Local File Read/Path Traversal bypass A hash fragment in an artifact URL slips past the earlier path check, letting the server resolve and read arbitrary local files.
- lobe-chat `/api/proxy` endpoint Server-Side Request Forgery vulnerability An unauthenticated proxy endpoint forwards attacker-supplied requests to arbitrary destinations, so internal services can be reached and their responses returned.
- Gradio's Component Server does not properly consider` _is_server_fn` for functions The Component Server does not check the server-function flag on functions, so non-server functions can be invoked.
- Gradio allows credential leakage on Windows Gradio releases before 4.20 leak credentials when running on Windows, with the advisory giving no further mechanism detail.
- sagemaker-python-sdk Command Injection vulnerability A requirements path passed to a dependency capture helper is inserted into a shell command without escaping, allowing injected commands.
- Flowise vulnerable to code injection via api/v1 A crafted script sent to the v1 API endpoint is executed by the Flowise server, giving a remote attacker code execution.
- PyTorch heap buffer overflow vulnerability A heap buffer overflow in PyTorch's vararg function handling can be triggered by crafted input, crashing the process.
- Pytorch use-after-free vulnerability A use-after-free in the mobile interpreter's runtime code lets memory be accessed after it has been released.
- Insecure deserialization in BentoML A crafted serialized object posted to any BentoML endpoint is deserialized by the server, running attacker-chosen OS commands.
- gradio vulnerable to Path Traversal The component endpoint lets callers invoke a block method that copies arbitrary host files into a cache directory, which they can then download.
- gradio Server-Side Request Forgery vulnerability A file parameter in a GET request makes the server contact internal addresses, so the responses reveal which internal ports are open.
- langchain vulnerable to path traversal Path traversal sequences in keys passed to LocalFileStore escape its root directory, letting callers read or write arbitrary files.
- Cross-site Scripting (XSS) in mindsdb/mindsdb A CSV uploaded with embedded JavaScript is rendered without escaping when viewed, so the script runs in the viewer's browser.
- mlflow vulnerable to Path Traversal Double URL decoding in artifact deletion lets encoded traversal sequences escape path validation and delete arbitrary server directories.
- mlflow vulnerable to Path Traversal Semicolons in request URLs smuggle path traversal sequences into the params portion, letting attackers reach files outside the intended directory.
- mlflow Path Traversal vulnerability Crafted artifact_location and source values using a '#' local URI bypass insufficient path validation, letting requests read arbitrary server files.
- mlflow vulnerable to Path Traversal Local-path detection mishandles empty or file-scheme URIs, so crafted model version sources are treated as non-local and can read arbitrary files.
- mlflow vulnerable to Path Traversal A crafted source parameter slips past the relative-path check during model version creation, letting attackers read arbitrary server files via the artifact handler.
- mlflow vulnerable to Path Traversal A fragment character in the artifact location makes the server resolve paths past checks meant to stop traversal, reading arbitrary files.
- Directory traversal in zenml The steps endpoint's logs URI accepts traversal sequences, letting callers read files outside the intended directory.
- zenml Session Fixation vulnerability JWTs remain valid after logout, so a captured token can still be replayed to access the victim's session.
- Ollama DNS rebinding vulnerability DNS rebinding lets a malicious page reach the localhost-only API, exposing chat, model deletion and resource exhaustion remotely.
- gradio Server-Side Request Forgery vulnerability An X-Direct-Url header adds attacker-chosen addresses to the replica list used by the proxy route, so requests are forwarded to internal endpoints.
- PaddlePaddle allows arbitrary file read via paddle.vision.ops.read_file The read_file operation opens arbitrary filesystem paths supplied to it, letting a caller read any file the process can access.
- PaddlePaddle Path Traversal vulnerability Path traversal sequences in a file path let attackers overwrite files outside the intended directory.
- PaddlePaddle command injection vulnerability Input passed to IrGraph.draw reaches a shell command, letting an attacker inject operating system commands.
- PaddlePaddle vulnerable to remote code execution PaddlePaddle 2.6.0 allows remote code execution from an input the advisory does not specify.
- PaddlePaddle command injection in paddle.utils.download._wget_download Input that gets past a filter reaches a wget download helper that builds a shell command, letting injected commands run.
- LangChain directory traversal vulnerability A ../ sequence in the load_chain path escapes the hub-only directory, letting attackers load configuration from arbitrary locations.
- ZenML Server Remote Privilege Escalation Vulnerability The user activation endpoint accepts a username plus a new password without proper authorisation, letting a remote caller take over accounts.
- MLFlow Cross-site Scripting vulnerability leads to client-side Remote Code Execution Dataset table fields are rendered without sanitization, so an untrusted dataset injects script into the notebook view and runs code on the client.
- Onnx Out-of-bounds Read vulnerability An off-by-one in the assertion macros' string copy causes an out-of-bounds read.
- Gradio apps vulnerable to timing attacks to guess password Password comparison returns at the first mismatched character, so response timing leaks progress and enables brute force because there is no rate limiting.
- Label Studio vulnerable to Cross-site Scripting if `<Choices>` or `<Labels>` are used in labeling config Imported task data is rendered unsanitised inside Choices or Labels tags, so an embedded image handler runs script in other users' browsers.
- Allegro AI ClearML Stores Credentials in Plaintext in MongoDB Instance ClearML stores user passwords in plaintext in its MongoDB backend, so anyone who reads the database obtains working credentials.
- Allegro AI ClearML path traversal vulnerability A dataset with crafted file paths makes the ClearML client SDK write files to arbitrary locations on the user's system.
- Gradio Path Traversal vulnerability A crafted JSON value in a Gradio API request makes the server include and read an arbitrary local file.
- Label Studio SSRF on Import Bypassing `SSRF_PROTECTION_ENABLED` Protections The URL import validates the target with one DNS lookup, so redirects or DNS rebinding let requests reach internal hosts despite the SSRF setting.
- Cross-site Scripting Vulnerability on Data Import Remote import saves downloaded content under a URL-derived name, so an attacker-hosted HTML file runs script when viewed on the site.
- Cross-site Scripting Vulnerability on Avatar Upload Avatar upload only checks that a file parses as an image, so a crafted file containing HTML is served and rendered, running script in the site origin.
- ReDoS in Embedchain A long crafted string sent to the JSON loader triggers catastrophic regex backtracking, tying up CPU for the duration of parsing.
- Code execution in Embedchain A crafted OpenAPI file parsed by the Embedchain loader reaches unsafe yaml.load, executing code on the server.
- Code Injection in paddlepaddle A URL parameter is concatenated into a shell command without validation, so a malicious URL can inject additional commands.
- PaddlePaddle floating point exception in paddle.argmin and paddle.argmax Crafted inputs to paddle.argmin and paddle.argmax trigger a floating point exception that crashes the runtime.
- PaddlePaddle segfault in paddle.put_along_axis A null pointer dereference in a tensor function crashes the process when given specific inputs.
- PaddlePaddle command injection in convert_shape_compare Attacker-supplied input reaching the convert_shape_compare routine is interpolated into a system command, so the attacker runs arbitrary OS commands on the host.
- PaddlePaddle stack overflow in paddle.searchsorted Crafted input to the searchsorted tensor operation overflows the stack, causing a crash or corrupting memory.
- PaddlePaddle null pointer dereference in paddle.nextafter A null pointer dereference in paddle.nextafter crashes the runtime when it is called.
- PaddlePaddle heap buffer overflow in paddle.repeat_interleave A heap buffer overflow in the repeat_interleave tensor operation lets crafted arguments write past allocated memory.