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
- Self-hosted n8n has Legacy Code node that enables arbitrary file read/write In legacy Code node mode, internal helper functions remain reachable from workflow scripts, letting editors read and write host files outside the intended restrictions.
- LangChain serialization injection vulnerability enables secret extraction in dumps/loads APIs Serialization does not escape dictionaries with the reserved lc key, so attacker-supplied data is later rebuilt as framework objects, which can expose secrets.
- n8n Vulnerable to Remote Code Execution via Expression Injection Workflow expressions from authenticated users are evaluated in an insufficiently isolated context, letting them run code as the n8n process.
- Langflow vulnerable to Server-Side Request Forgery The API Request node fetches any user-supplied URL without blocking internal or metadata addresses and returns the response to a caller holding only an API key.
- External Control of File Name or Path in Langflow An authenticated user can set an arbitrary fs_path in a flow request, so the server writes the flow JSON to any file path.
- Ollama Platform has missing authentication enabling attackers to perform model management operations Several Ollama API endpoints accept requests without authentication, so remote callers can perform model management operations.
- Weaviate OSS has a Path Traversal Vulnerability via Backup ZipSlip Backup archive entries with absolute or parent-directory paths let restore write files outside the restore root.
- Weaviate OSS has path traversal vulnerability via the Shard Movement API The fileName value in the shard transfer logic is not validated, letting a caller read arbitrary files the service process can access.
- LangGraph's SQLite is vulnerable to SQL injection via metadata filter key in SQLite checkpointer list method Metadata filter keys are concatenated into SQL in the SQLite checkpointer's list method, so a crafted key injects arbitrary query logic.
- n8n vulnerable to Remote Code Execution via Git Node Custom Pre-Commit Hook A workflow can point Git's hooks path at a repository directory containing a pre-commit script, which Git then runs as a local command on the n8n host.
- Langflow CORS misconfiguration enables Account Takeover and RCE Permissive CORS with credentials and a cross-site refresh cookie let a malicious page obtain fresh tokens for a logged-in victim.
- Open WebUI vulnerable to Server-Side Request Forgery (SSRF) via Arbitrary URL Processing in /api/v1/retrieval/process/web An authenticated user can make the retrieval web-processing endpoint fetch any URL because the submitted address is not validated.
- open-webui is Vulnerable to Incorrect Access Control The task-stop endpoint cancels any LLM task by ID without checking that the requesting user owns it.
- Open WebUI Vulnerable to Stored DOM XSS via Note 'Download PDF' SVG markup in imported Markdown is inserted into the PDF export via innerHTML without sanitising, so script runs in the exporting user's session.
- Keras Directory Traversal Vulnerability A crafted tar archive uses symlink path handling to evade Keras's safe-path filter, so extraction writes files outside the target directory.
- Ray's New Token Authentication is Disabled By Default Token authentication is off by default on Ray management interfaces, so a network attacker can submit jobs that run arbitrary code.
- Ray is vulnerable to Critical RCE via Safari & Firefox Browsers through DNS Rebinding Attack A malicious webpage uses DNS rebinding and a spoofed User-Agent to reach unauthenticated Ray job endpoints on a developer's machine and run code.
- new-api is vulnerable to SSRF Bypass A redirect from an allowed external URL points the server at internal addresses, because the SSRF check only validates the first request.
- MLX has heap-buffer-overflow in load() A crafted NumPy .npy header makes the loader read past a heap buffer because it truncates the header at a null byte and skips a length check.
- LangChain Vulnerable to Template Injection via Attribute Access in Prompt Templates Untrusted template strings let attackers traverse object attributes and dictionary keys in prompt templates, reaching secrets such as environment variables.
- Flowise has Authentication Bypass Using Unprotected Registration Endpoint (/register) Flowise's registration endpoint is open by default, so anyone can create an account and log in.
- Flowise does not Prevent Bypass of Password Confirmation - Unverified Password Change Authenticated users can change their password without supplying the current one, so a stolen session can take over the account.
- Flowise doesn't Prevent Bypass of Password Confirmation through Unverified Email Change (credentials) Changing the account email takes effect without verification, so an attacker can swap in a recovery address and take over the account.
- Flowise Fails to Invalidate Existing Sessions After Password Changes Existing sessions remain valid after a password change, so an attacker holding a stolen session token keeps access.
- Milvus Proxy has a Critical Authentication Bypass Vulnerability Milvus Proxy trusts a client-supplied sourceID header, so unauthenticated requests can skip authentication and gain administrative access.
- Open WebUI vulnerable to Stored DOM XSS via prompts when 'Insert Prompt as Rich Text' is enabled resulting in ATO/RCE Prompt bodies inserted as rich text are converted to HTML and assigned to innerHTML without sanitisation, so stored markup runs in other users' browsers.
- Vercel’s AI SDK's filetype whitelists can be bypassed when uploading files The AI SDK's file upload type allowlist can be circumvented, letting files of disallowed types be accepted.
- Agno session state overwrites between different sessions/users A race condition under concurrent load can save one session's state into another session, exposing a user's data to a different user.
- n8n Vulnerable to Remote Code Execution via Git Node Pre-Commit Hook Committing from a workflow's Git node after cloning an untrusted repository runs that repository's pre-commit hook, executing attacker code in n8n.
- LangGraph SQLite Checkpoint Filter Key SQL Injection POC for SqliteStore Filter keys taken from user input are concatenated into a JSON path inside a SQL query without escaping, allowing injected SQL.
- MLflow Tracking Server Model Creation Directory Traversal Remote Code Execution Vulnerability A model-creation request supplies an unvalidated path that allows traversal to files that lead to code execution as the service account, without authentication.
- MLflow Weak Password Requirements Authentication Bypass Vulnerability Weak password requirements let remote, unauthenticated attackers get past MLflow's authentication.
- LangGraph's SQLite store implementation has a SQL Injection Vulnerability Filter operator handling concatenates input into SQL in the SQLite checkpoint store, so crafted filters can read other records and stored secrets.
- Lobe Chat vulnerable to Server-Side Request Forgery with native web fetch module The crawlPages endpoint fetches client-supplied URLs server-side without filtering internal or metadata addresses, letting callers reach internal services.
- llama-index has Insecure Temporary File A shared, world-writable NLTK data directory lets local users replace or corrupt files that the library later loads.
- FlowiseAI/Flosise has File Upload vulnerability The upload endpoint accepts any file type and content, letting an authenticated user plant a server script that can run commands.
- vLLM is vulnerable to Server-Side Request Forgery (SSRF) through `MediaConnector` class The media loader fetches any http, https or file URL supplied in a multimodal request without host restrictions, so the server can be made to reach internal addresses.
- vLLM is vulnerable to timing attack at bearer auth The API key is checked with a non-constant-time string comparison, so response timing reveals correct characters and the key can be recovered incrementally.
- LangChain Text Splitters is vulnerable to XML External Entity (XXE) attacks due to unsafe XSLT parsing Developer-supplied XSLT stylesheets are parsed without hardening, so external entities or document() calls can read local files or fetch URLs.
- clearml is vulnerable to Path Traversal through its `safe_extract` function The safe archive extraction routine mishandles symbolic and hard links, letting archive entries write outside the target directory.
- Flowise is vulnerable to stored XSS via "View Messages" allows credential theft in FlowiseAI admin panel Chat message content is rendered unsanitised in the admin message viewer, so injected script runs for the admin and can read localStorage.
- lobe-chat has an Open Redirect The OIDC redirect builds its target host from client-supplied forwarded headers, so an attacker can send users to an arbitrary domain.
- Hugging Face Transformers vulnerable to Regular Expression Denial of Service (ReDoS) in the AdamWeightDecay optimizer Attacker-supplied regex patterns in the optimizer's weight-decay filters trigger catastrophic backtracking during matching, hanging the training process.
- H2O affected by a deserialization vulnerability JDBC connection parameter handling can be tricked past regex checks using double URL encoding, enabling file reads and code execution.
- InvokeAI has External Control of File Name or Path The bulk download endpoint joins a user-supplied item name into a filesystem path without canonicalisation, so unauthenticated requests can read files outside the media directory.
- FlowiseAI Pre-Auth Arbitrary Code Execution A filter expression string is compiled into a JavaScript function and executed on the server, letting the filter run arbitrary operating system commands.
- Flowise has an Arbitrary File Read An unauthenticated request sets the chatId parameter to a traversal path, which is joined onto the storage root because only chatflowId is validated, letting the attacker read arbitrary local files.
- FlowiseAI/Flowise has Server-Side Request Forgery (SSRF) vulnerability The link-extraction endpoint fetches any caller-supplied URL without validation, so callers can make the server request internal hosts and parse the responses.
- Flowise has arbitrary file access due to missing chat flow id validation The chatflowId from the request body is joined into a storage path without validation, so traversal sequences read or write files outside the storage directory.
- Stored XSS in n8n LangChain Chat Trigger Node via initialMessages Parameter A script stored in the initial messages setting runs in the browser of every visitor to the public chat page.