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
- Gradio vulnerable to arbitrary file read with File and UploadButton components Omitting upload metadata in a File component payload lets a caller supply a server file path and read arbitrary files.
- LocalAI Cross-site Scripting vulnerability Crafted parameters sent to the delete model API store a script that runs in the browser of users who open the homepage.
- AgentScope uses `eval` A helper evaluates a user-supplied string with eval, so the submitted expression runs arbitrary code on the host.
- gradio Server Side Request Forgery vulnerability The save_url_to_cache function fetches any supplied URL without restriction, letting an attacker make the server download local or internal resources.
- Lollms vulnerable to Cross-site Scripting The SVG sanitiser removes only script tags and on-event attributes, so other SVG vectors survive and run script when a victim opens the file.
- Langchain Path Traversal vulnerability A path helper does not sanitise user input, letting callers write, overwrite, read and delete files outside the intended directory.
- Lord of Large Language Models (LoLLMs) path traversal vulnerability in the api open_personality_folder endpoint The personality folder parameter bypasses sanitize_path through traversal sequences, letting callers open arbitrary host directories.
- Lord of Large Language Models (LoLLMs) Server path traversal vulnerability in lollms_file_system.py RAG database endpoints skip path sanitisation, so a caller can vectorise SQLite files from any directory on the host.
- Gradio's dropdown component pre-process step does not limit the values to those in the dropdown list Crafted requests can send values outside the dropdown's configured list, ignoring the allow_custom_value restriction on the server side.
- Gradio uses insecure communication between the FRP client and server The tunnel used by share=True runs without enforced HTTPS, so an on-path observer can read uploaded files and modify traffic.
- Gradio has a one-level read path traversal in `/custom_component` A path parameter on the custom component endpoint can climb one directory level, letting remote requests read source files of custom components.
- Gradios's CORS origin validation is not performed when the request has a cookie Gradio skips CORS origin validation for requests carrying cookies, so a malicious site can make authenticated cross-origin requests to a local server.
- Gradio has several components with post-process steps allow arbitrary file leaks Gradio file components accept arbitrary file paths from requests that bypass intended input constraints, so the server returns sensitive local files.
- Gradio vulnerable to SSRF in the path parameter of /queue/join The queue join endpoint makes the server fetch attacker-supplied URLs and cache the results, reaching internal hosts.
- Gradio's `is_in_or_equal` function may be bypassed Path checks that rely on parent-directory sequences fail to normalise certain inputs, letting requests reach files outside the permitted directory.
- Gradio's CORS origin validation accepts the null origin The localhost origin list includes the null origin, so sandboxed iframes can make cross-origin requests and read the responses.
- Gradio has an XSS on every Gradio server via upload of HTML files, JS files, or SVG files Users who upload HTML, JavaScript or SVG files to Gradio can have scripts run in other users' browsers when those files are viewed.
- In Gradio, the `enable_monitoring` flag set to `False` does not disable monitoring Setting enable_monitoring to False does not block the /monitoring endpoint, so anyone who requests it can see application analytics.
- Gradio performs a non-constant-time comparison when comparing hashes The analytics dashboard compares secret hashes in non-constant time, so response timing lets an attacker recover the hash byte by byte.
- Gradio has a race condition in update_root_in_config may redirect user traffic A race condition lets an attacker change the root URL the Gradio frontend uses, redirecting user requests and data to a malicious server.
- open-webui allows writing and deleting arbitrary files The pipelines upload endpoint joins an unsanitised client filename onto the cache directory, allowing files outside it to be written or deleted.
- open-webui Insecure Direct Object Reference (IDOR) vulnerability The memory update endpoint checks only that a memory ID exists, not that it belongs to the caller, so users can rewrite others' stored memories.
- DeepSpeed Remote Code Execution Vulnerability A command injection flaw in DeepSpeed lets an attacker pass crafted input that the library executes as shell commands on the host.
- Inefficient Regular Expression Complexity in langflow A regex with catastrophic backtracking runs on attacker-supplied text in an HTTP handler, tying up the server.
- Gradio allows users to access arbitrary files Modifying requests to a Gradio app's server lets a remote user read arbitrary files from the host machine.
- Flowise and Flowise Chat Embed vulnerable to Stored Cross-site Scripting The chat embed does not sanitise stored input, so injected script runs in viewers' browsers when messages are displayed.
- lobe-chat implemented an insufficient fix for GHSA-mxhq-xw3g-rphc (CVE-2024-32964) The proxy's SSRF check validates only the initial URL and ignores redirects, so an external URL can redirect the server to private or loopback addresses.
- vLLM denial of service vulnerability A completions API request with an empty prompt crashes the serving process.
- Composio Path Traversal vulnerability The file parameter in Composio's API server is not restricted to an intended directory, so crafted values can read or write files outside it.
- LiteLLM Server-Side Request Forgery (SSRF) vulnerability A user-supplied api_base on the chat completions route makes the server send requests to attacker hosts, carrying the stored OpenAI key with them.
- MindsDB Cross-site Scripting vulnerability Script stored in engine, database, project or dataset names runs in the browser of any user who lists them in the web UI.
- MindsDB Eval Injection vulnerability A crafted INSERT query against a ChromaDB-backed database is passed to Python eval, running attacker code on the server.
- MindsDB Eval Injection vulnerability A crafted INSERT into a SharePoint-backed table passes attacker-supplied Python to eval on the server.
- MindsDB Eval Injection vulnerability An UPDATE query carrying Python code is passed to eval on the server when certain integrations are installed.
- MindsDB Eval Injection vulnerability A crafted INSERT against a SharePoint-backed database puts Python code into a value that the server passes to eval.
- MindsDB Eval Injection vulnerability Python code embedded in a crafted SELECT WHERE clause is passed to eval when run against a Weaviate-backed database, executing on the server.
- MindsDB Eval Injection vulnerability Python code inside an INSERT query for the SharePoint engine is passed to eval and executed on the MindsDB server.
- H2O.ai H2O vulnerable to deserialization attacks via a JDBC Connection URL The import endpoint accepts a caller-chosen JDBC URL, and crafted driver parameters trigger deserialization, file reads and command execution on the server.
- MindsDB Vulnerable to Bypass of SSRF Protection with DNS Rebinding The private-address check resolves the hostname once, so a DNS record that later returns an internal address redirects the real request to internal services.
- Ollama can extract members of a ZIP archive outside of the parent directory Zip archive entries containing parent-directory path segments are written outside the intended extraction folder.
- Flowise Authentication Bypass vulnerability Unauthenticated remote requests reach administrator-level API endpoints because access checks are missing in this version.
- Flowise Unauthenticated Denial of Service (DoS) vulnerability An unauthenticated request to the upload file retrieval endpoint with crafted input crashes the whole Flowise instance.
- LlamaIndex includes an exec call for `import {cls_name}` A class name is interpolated into an exec() statement in the download integration code, so a crafted name runs arbitrary Python.
- Khoj Vulnerable to Stored Cross-site Scripting In Automate (Preview feature) The automation instruction parameter is rendered into the config page without escaping, so injected HTML and script run in viewers' browsers.
- Flowise Cross-site Scripting in api/v1/chatflows/id An unknown chatflow ID is echoed unescaped into an HTML 404 page, so a crafted URL runs script in the victim's browser.
- Flowise Cors Misconfiguration in packages/server/src/index.ts Flowise sends a wildcard CORS header, letting any origin read responses, and chained path injection can expose arbitrary server files.
- Flowise Cross-site Scripting in /api/v1/chatflows-streaming/id An unknown chatflow ID is echoed unescaped into a text/html 404 page, letting a crafted URL inject script.
- Flowise Cross-site Scripting in /api/v1/public-chatflows/id The 404 page reflects the unvalidated chatflow ID as text/html, so injected script runs in users' sessions.
- Flowise Path Injection at /api/v1/openai-assistants-file The file endpoint uses an unsanitised fileName parameter, letting callers read arbitrary files on the server.
- Flowise Cross-site Scripting in/api/v1/credentials/id An unknown ID is reflected unescaped into a text/html 404 page, letting attacker-supplied script run in a victim's session.