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
- NextChat has full-read SSRF and XSS vulnerability in /api/cors endpoint An open CORS proxy endpoint fetches arbitrary URLs with any HTTP method and returns the content, reaching internal services and enabling script injection.
- Insecure Jinja2 templates rendered in Haystack Components can lead to RCE Jinja2 templates that users can supply to Haystack components are rendered in a way that runs arbitrary code on the client machine.
- TensorFlow has segfault in array_ops.upper_bound Passing a tensor that is not rank 2 to upper_bound makes the op read invalid memory and crash with a segmentation fault.
- Guardrails AI vulnerable to Improper Restriction of XML External Entity Reference Guardrails parses attacker-supplied RAIL XML with external entities enabled, so a SYSTEM entity can read local files.
- LoLLMS vulnerable to Expected Behavior Violation The discussion database name passes through an inadequate path sanitiser, letting crafted values point writes at locations outside the intended directory.
- TorchServe gRPC Port Exposure TorchServe binds its gRPC management and inference ports to all interfaces by default, exposing them to network clients without local restriction.
- TorchServe vulnerable to bypass of allowed_urls configuration Dot-segment URLs slip past the allowed_urls check, letting an attacker download a model from disallowed locations into the model store.
- Vanna vulnerable to SQL Injection Unauthenticated requests reach SQL built from user input in the Flask API, letting an attacker call file-reading database functions to read local files.
- OpenAI's ChatGPT Mac App Stored User Data in Unencrypted Text Files The ChatGPT macOS app stored user conversations as unencrypted plain text, so anyone with local file access could read them; a third party demonstrated it and OpenAI fixed it.
- Reflected Cross-Site Scripting (XSS) in zenml An unvalidated redirect parameter is reflected into the survey page, letting a crafted link run script in a victim's session.
- h2o vulnerable to unexpected POST request shutting down server The rapids run_tool endpoint can invoke the main method of arbitrary tool classes, and an invalid argument to one of them crashes the server.
- pytorch-lightning vulnerable to Arbitrary File Write via /v1/runs API endpoint Plugin tar.gz archives containing ../ entries are extracted without sanitising names, so an attacker can write files anywhere the service can reach.
- litellm vulnerable to remote code execution based on using eval unsafely A crafted base64 payload sent to the config update endpoint is decoded and passed to eval when secrets are fetched, running code on the server.
- litellm vulnerable to improper access control in team management Team management endpoints skip authorisation checks, so any caller can view, change or delete teams and their members.
- lollms vulnerable to path traversal due to unauthenticated root folder settings change Unauthenticated settings changes set the XTTS root folder to the filesystem root, letting the file read endpoint serve arbitrary files and redirecting audio output elsewhere.
- lollms path traversal vulnerability allows overriding of config.yaml file, leading to RCE A personality config endpoint accepts traversal in its file path, letting callers overwrite the server config and turn on code-related settings.
- lollms vulnerable to dot-dot-slash path traversal in XTTS server The tts_to_file endpoint accepts unvalidated user-supplied paths, so dot-dot sequences let requests write audio files to arbitrary locations.
- Remote Code Execution in create_conda_env function in lollms The environment name and Python version parameters are interpolated into a shell command run with shell=True, letting an attacker append arbitrary commands.
- Improper line feed handling in zenml A newline character in component names submitted through the API makes the server consume excessive resources, breaking component and stack creation.
- Open redirect in gradio Improper validation of user-supplied URLs lets attackers send victims through redirects to arbitrary external websites.
- Remote Code Execution via path traversal bypass in lollms Empty category and folder values let the extension path resolve to the root, so a planted config.yaml causes an __init__.py to run.
- parisneo/lollms Local File Inclusion (LFI) attack The path sanitiser does not treat Windows backslash separators as traversal, so crafted paths reach file reads and deletes through endpoints such as personalities and preset deletion.
- Langflow remote code execution vulnerability Users who can reach the custom component endpoint can submit Python code that the server executes.
- zenml-io/zenml does not expire the session after password reset Old session tokens remain valid after a password change, so a previously stolen session keeps working in another browser.
- Server-Side Request Forgery in gradio Gradio's queue endpoint fetches attacker-supplied URLs server-side through a cache helper without validation, reaching internal hosts and cloud metadata services.
- Arbitrary system path lookup in h20 The typeahead lookup accepts a root path and returns filesystem directory listings to remote users without restriction.
- Denial of service in langchain-community A sitemap that references itself makes parse_sitemap recurse without limit, exceeding Python's recursion depth and crashing the process.
- Remote code execution in pytorch lightning A crafted serialized delta passes the deserializer whitelist while setting dunder attributes, reaching modules that let the attacker execute code on the server.
- Arbitrary file deletion in litellm The transcription endpoint deletes the path taken from the uploaded filename without validation, so callers can remove arbitrary files on the server.
- SQL injection in litellm The team update route places the user_id value directly into a raw SQL delete statement, letting attackers run blind queries to extract stored keys and tokens.
- SQL injection in litellm The /global/spend/logs endpoint concatenates the api_key query parameter directly into a SQL statement instead of using parameters.
- LoLLMS Path Traversal vulnerability Weak Windows path sanitisation lets crafted input escape the intended directory and read arbitrary files.
- LoLLMS Path Traversal vulnerability Backslash-separated paths sent to the user_infos endpoint escape the intended directory on Windows, letting callers read or delete arbitrary files.
- Remote code execution in mlflow Filenames taken from a remote HTTP response or URL are joined into local paths unsanitised, allowing traversal or absolute-path writes that lead to command execution.
- Undefined Behavior in mlflow URL-encoded and decoded model names are treated as different models, letting an attacker register a same-named model that users later load by mistake.
- Local File Inclusion in mlflow Directory traversal sequences placed in a URI fragment evade the validation added earlier for query strings, letting requests read arbitrary local files.
- Improper authorization in zenml Any authenticated user can update other accounts via the PUT users endpoint, including setting active to false to deactivate them.
- Race condition in zenml Parallel user creation requests with the same username both succeed, producing duplicate accounts and inconsistent user records.
- Improper authentication in zenml The password change endpoint skips verification of the current password, so anyone with an active session can take over the account.
- Clickjacking in zenml Missing X-Frame-Options and Content-Security-Policy headers let attackers embed the ZenML UI in a malicious iframe and trick users into unintended clicks.
- Cross site scripting in zenml A stored script payload placed in the logo URL field runs in other users' browsers when the page renders it.
- Arbitrary JavaScript execution due to using outdated libraries A crafted PDF embeds script in font data that an outdated pdf.js build executes in the viewer when the app renders it.
- qdrant input validation failure A crafted snapshot archive containing symlinks lets recovery read and write files outside the intended data directory.
- path traversal vulnerability was identified in the parisneo/lollms-webui A crafted category parameter in the personalities listing endpoint escapes the intended folder and reveals subfolder names of arbitrary directories.
- code injection vulnerability exists in the huggingface/text-generation-inference repository A malicious fork branch name is interpolated into a GitHub Actions workflow command, so the branch name runs shell code on the CI runner.
- Ollama does not validate the format of the digest (sha256 with 64 hex digits) Model blob lookups accept digests without checking their format, so crafted strings containing ../ resolve to file paths outside the blob store.
- Reportedly Hacked AI-Powered Robot Vacuums Allegedly Used for Surveillance and Harassment Hackers exploited a flaw in Ecovacs Deebot X2 robot vacuums to gain unauthorized access to their cameras and microphones, enabling privacy invasion and harassment.
- Gradio applications running locally vulnerable to 3rd party websites accessing routes and uploading files A locally running Gradio server accepts requests from arbitrary websites, so a visited page can upload files to the user's machine.
- litellm passes untrusted data to `eval` function without sanitization Values written to environment variables through a config update reach an unsafe eval() call during secret retrieval with Google KMS, executing code.
- LoLLMS Command Injection vulnerability The unInstall_binding endpoint uses an unsanitised name that traverses directories, letting an attacker load a malicious __init__.py and run code.