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
- n8n: LDAP Email-Based Account Linking Allows Privilege Escalation and Account Takeover LDAP login automatically links an identity to any local account with a matching email, so a user who controls their LDAP email can take over an admin account.
- n8n Has External Secrets Authorization Bypass in Credential Saving Saving a credential that references an external secret by name returns its plaintext value without the external-secret list permission being checked.
- n8n is Vulnerable to Credential Theft via Name-Based Resolution and Permission Checker Bypass in Community Edition A member user resolves another user's credential by ID, and a permission check skips generic HTTP credential types, letting a workflow decrypt and use them.
- n8n Has Authorization Bypass in OAuth Callback via N8N_SKIP_AUTH_ON_OAUTH_CALLBACK With an environment flag set, the OAuth callback skips verifying that the state belongs to the current user, so tokens are stored in an attacker-controlled credential.
- n8n has In-Process Memory Disclosure in its Task Runner An authenticated user can allocate uninitialised buffers in the task runner, exposing leftover in-process memory such as tokens and secrets.
- New API has passkey-based secure step-up verification bypass for root-only channel secret disclosure The secure verification endpoint marks a passkey step-up as complete when the account merely has a passkey registered, without checking an actual assertion.
- New API: IDOR in VideoProxy allows cross-user video content access via missing ownership check The video proxy looks up generated tasks by ID alone without checking ownership, so any authenticated user can fetch another user's video.
- langflow has Unauthenticated IDOR on Image Downloads The image download route omits the authentication and ownership checks used by other file routes, so anyone with a flow ID and filename can fetch images.
- langflow: /profile_pictures/{folder_name}/{file_name} endpoint file reading The profile picture endpoint concatenates path parameters without checks, so traversal sequences let unauthenticated callers read arbitrary server files.
- Langflow has an Arbitrary File Write (RCE) via v2 API The v2 upload route takes the multipart filename unchecked and writes via naive path joining, so authenticated users can write outside the upload directory.
- NLTK has a Downloader Path Traversal Vulnerability (AFO) - Arbitrary File Overwrite The NLTK downloader trusts subdir and id values from remote XML indexes, so ../ sequences write files and directories outside the download folder.
- Unauthenticated remote shutdown in nltk.app.wordnet_app Any client that can reach the WordNet browser's port can stop the server with one unauthenticated GET request.
- Langflow is Missing Ownership Verification in API Key Deletion (IDOR) The API key delete route checks only that the caller is logged in, then deletes any key by ID without confirming ownership.
- Improper Neutralization of Input During Web Page Generation ('Cross-site Scripting') in nltk A crafted lookup URL places its word into the WordNet browser page unescaped, so injected markup runs in the local app's origin.
- Natural Language Toolkit (NLTK) has unbounded recursion in JSONTaggedDecoder.decode_obj() may cause DoS The JSON tagged decoder recurses without a depth limit, so deeply nested input exceeds the recursion limit and crashes the Python process.
- Unauthenticated Remote Code Execution in Langflow via Public Flow Build Endpoint An unauthenticated public build endpoint accepts attacker-supplied flow data whose node definitions contain Python that the server runs with exec.
- Ray Dashboard is vulnerable to path traversal through its static file handling mechanism Traversal sequences in static file requests escape the dashboard's static directory, exposing arbitrary local files to remote callers.
- MLflow has a command injection in mlflow/sagemaker/__init__.py The sagemaker CLI interpolates the container image name into a shell command run through os.system, so shell metacharacters in the flag execute.
- SGLang's multimodal generation module is vulnerable to unauthenticated remote code execution through the ZMQ broker The ZMQ broker in SGLang's multimodal module unpickles network messages without authentication, so a remote sender can run code.
- Flowise affected by Server-Side Request Forgery (SSRF) in HTTP Node Leading to Internal Network Access The HTTP node in Flowise sends server-side requests to user-supplied URLs with no default restriction on private, loopback or metadata addresses.
- NLTK has Arbitrary File Read via Absolute Path Input in nltk.util.filestring() A utility function opens whatever path it is given, so absolute or traversal paths let callers read arbitrary local files.
- vLLM has SSRF Protection Bypass vLLM's URL check parses backslash-containing URLs differently from the aiohttp client that fetches them, letting a request reach an attacker host.
- Flowise Missing Authentication on NVIDIA NIM Endpoints The NVIDIA NIM route prefix is whitelisted past global auth, so unauthenticated callers reach token retrieval and container management endpoints.
- Flowise has IDOR leading to Account Takeover and Enterprise Feature Bypass via SSO Configuration The login-method update endpoint trusts an organization ID from the request body without checking the caller belongs to that organization, so any user can overwrite its SSO settings.
- Flowise has Arbitrary File Upload via MIME Spoofing The unauthenticated upload endpoint trusts the client-supplied Content-Type, so attackers can store active files such as scripts by spoofing an allowed type.
- Flowise Allows Mass Assignment in `/api/v1/leads` Endpoint Object.assign copies every field from the unauthenticated request body onto the Lead entity, letting callers set internal fields such as id and chatId.
- Flowise has Authorization Bypass via Spoofed x-request-from Header The API middleware trusts a client-supplied internal header to skip API key checks, letting any authenticated tenant session reach administrative endpoints.
- RAGAS has an Arbitrary File Read vulnerability Ragas' multimodal handling accepts URLs in retrieved context without adequate validation, so a crafted URL can make it read local files.
- SageMaker Python SDK replaced eval() with safe parser in JumpStart search functionality The search_hub query parameter is passed to eval(), so crafted input in the JumpStart search path can run arbitrary Python.
- Flowise Vulnerable to PII Disclosure on Unauthenticated Forgot Password Endpoint The forgot-password endpoint returns the full user record, including email and identifiers, to unauthenticated callers, which reveals whether accounts exist.
- Flowise has Insufficient Password Salt Rounds A default bcrypt cost of 5 makes stored password hashes cheap to brute-force if the database is leaked.
- NLTK has a Path Traversal issue Corpus reader classes accept unvalidated file paths, so crafted names can reach files outside the corpus directory.
- Gradio is Vulnerable to Absolute Path Traversal on Windows with Python 3.13+ On Windows with Python 3.13 or later, root-relative paths pass Gradio's safe-join check, letting unauthenticated requests read arbitrary files.
- Gradio has SSRF via Malicious `proxy_url` Injection in `gr.load()` Config Processing A malicious remote Space config adds its own proxy host to the allowlist, letting the victim's server reach internal addresses.
- Gradio has an Open Redirect in its OAuth Flow Logout and login callback endpoints redirect to an unvalidated target URL, letting attackers send users from a trusted domain to external sites.
- n8n has Webhook Forgery on Zendesk Trigger Node Requests to a Zendesk trigger webhook URL are accepted without verifying the HMAC signature, so anyone knowing the URL can start the workflow with forged data.
- n8n: SQL Injection in MySQL, PostgreSQL, and Microsoft SQL nodes Database workflow nodes do not escape table and column identifiers, so a crafted name in node configuration injects arbitrary SQL.
- n8n has an Authentication Bypass in its Chat Trigger Node The Chat Trigger's n8n user authentication can be circumvented, letting unauthenticated requests reach the webhook workflow.
- n8n: Webhook Forgery on Github Webhook Trigger The GitHub trigger never checks HMAC signatures, so anyone who knows the webhook URL can send forged events that start the workflow.
- n8n has an SSO Enforcement Bypass in its Self-Service Settings API An SSO-authenticated user can disable SSO enforcement on their own account through a settings endpoint, then set a local password and log in directly.
- LangChain Community: redirect chaining can lead to SSRF bypass via RecursiveUrlLoader The URL check in RecursiveUrlLoader runs only on the starting address, so automatic redirects can send the crawler to internal or metadata endpoints.
- n8n Vulnerable to Stored XSS via Various Nodes A workflow author embeds script in node-rendered pages such as forms and chat, which then runs for any visitor.
- n8n has Unauthenticated Expression Evaluation via Form Node Unauthenticated form submissions with a leading equals sign make n8n evaluate the field as an expression, which can be chained with a sandbox escape to run code.
- n8n has a Sandbox Escape in its JavaScript Task Runner A workflow author can break out of the JavaScript task runner sandbox and execute code on the host outside its boundary.
- n8n has Arbitrary File Read via Python Code Node Sandbox Escape The Python Code node sandbox fails to restrict certain built-in objects, letting a workflow author read host files or execute code outside the sandbox.
- n8n: Expression Sandbox Escape Leads to RCE Crafted expressions in workflow parameters escape the expression sandbox and run system commands on the host for users who can edit workflows.
- n8n has Potential Remote Code Execution via Merge Node The Merge node's SQL query mode lets an authenticated workflow author run code and write files on the n8n host.
- MindsDB: Path Traversal in /api/files Leading to Remote Code Execution A crafted filename in the file upload endpoint escapes the intended upload directory, and the traversal can be chained to achieve command execution.
- New API has an SQL LIKE Wildcard Injection DoS via Token Search Authenticated users send LIKE patterns with unescaped wildcards to the token search, forcing the database to match and return very large row sets.
- MLflow Use of Default Password Authentication Bypass Vulnerability Hard-coded default credentials in the MLflow basic auth file let remote callers authenticate and reach administrator-level access.