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
- Flowise: Unauthenticated Credential Abuse via Text-to-Speech Endpoint Allows Unauthorized Use of Private Chatflow TTS Credentials An unauthenticated caller can use a private chatflow's text-to-speech credentials by supplying its ID, because the route skips the public-access check.
- Flowise: RBAC Bypass Leading to Unauthorized Workspace Variables Disclosure A user denied variables view permission can run custom JavaScript that receives all workspace variables, including server environment values, because $vars is injected unconditionally.
- Flowise: SSRF Protection Bypass via IPv4-Mapped IPv6 Addresses Private IPv4 deny rules are skipped when an IPv4-mapped IPv6 address is used, so hostnames resolving to such addresses reach internal targets.
- Flowise: Cross-workspace credential IDOR in openai-assistants-vector-store A client-supplied credential ID is used without checking that it belongs to the caller's workspace, letting one tenant use another tenant's OpenAI key.
- Flowise: Missing Authorization on Execution Update Endpoint A PUT route on execution records skips the permission middleware that sibling routes use, letting any authenticated user change any execution.
- Flowise: Broken Access Control in Stripe Subscription Endpoints Allows Cross-Tenant Billing Manipulation Billing routes accept a client-supplied Stripe subscription ID without checking it belongs to the caller's organisation, enabling changes to other tenants.
- Flowise: `DELETE /api/v1/chatflows/:id` does not validate resource type, allowing `agentflows:delete` and `chatflows:delete` to delete each other’s flow type The delete route accepts either permission but never checks flow type, so a key scoped to one flow kind can delete the other kind.
- Flowise: Unauthenticated OAuth2 token refresh endpoint returns access tokens — enables token theft for any connected service An unauthenticated route returns freshly refreshed OAuth access tokens, so anyone holding a credential ID can take them and use the victim's connected services.
- Flowise: Unauthenticated OAuth2 Refresh Enables Non-Blind SSRF and Secret Exfiltration The OAuth2 refresh route is whitelisted from auth and sends server-side requests to a credential-supplied token URL without SSRF checks, returning the remote response to the caller.
- Flowise: Incomplete Credential Redaction Exposes Secrets via API The credential endpoint redacts only password-typed fields, so secrets stored in string fields such as URLs and keys are returned in plaintext.
- Flowise: Cross-Workspace OAuth2 Credential Metadata Leak Flowise OAuth2 endpoints look up credentials by ID alone, so authenticated users reach other workspaces' credentials and unauthenticated callers can overwrite tokens.
- Flowise: Missing authorization on `/api/v1/files` allows low-privileged API keys to list and delete files across workspaces within the same organization The files API checks only a feature flag, not per-permission authorisation, so any API key in an organisation can list and delete other workspaces' files.
- Natural Language Toolkit (NLTK): Path Traversal in NKJPCorpusReader leads to Arbitrary File Read and bypasses the nltk.pathsec sandbox (ENFORCE=True) The corpus reader joins attacker-supplied file IDs to the root without containment checks, reading files outside the corpus while skipping the library's path validation.
- Natural Language Toolkit (NLTK): ReDoS in NLTK ReviewsCorpusReader FEATURES regex A regex in the reviews corpus reader backtracks quadratically on a long line without brackets, so one crafted line stalls parsing.
- Natural Language Toolkit (NLTK): DNS-rebinding SSRF filter bypass in nltk.pathsec.urlopen (nltk.download / nltk.data.load) defeats ENFORCE mode The SSRF check resolves a hostname separately from the actual connection, so DNS rebinding returns a public address for validation and an internal one for connecting.
- Natural Language Toolkit (NLTK) has path traversal in FramenetCorpusReader.frame() that allows arbitrary XML file read, bypassing the nltk.pathsec sandbox (ENFORCE=True) A name containing parent-directory sequences is joined into a corpus XML path read with plain open, skipping the pathsec check and reading files outside the corpus.
- NLTK vulnerable to Eval Injection via collocations CLI arguments Command-line arguments are passed straight into eval as attribute names, letting an attacker inject Python expressions that run OS commands.
- Hugging Face Datasets folder-based builders allow path traversal through file_name metadata Folder-based dataset builders join an unvalidated file_name metadata value to the dataset directory, so traversal sequences read arbitrary local files.
- ONNX: Heap-Buffer-Overflow READ in Gemm Version Converter Adapter via Undersized Input Shape The Gemm 7-to-6 version adapter reads shape entries without checking tensor rank, so a model with a one-dimensional input reads past its heap allocation.
- Open WebUI: Cross-user model-list exposure via static cache key in get_all_models (aiocache key= vs key_builder= misuse) A misused cache key stores per-user permission-filtered model lists under one shared key, so a different user can receive them within the TTL.
- Open WebUI: Private channel messages can be disclosed through cross-channel thread parent_id binding A low-privilege user requests a thread in an accessible channel and supplies a message id from a private channel as the root, so the parent message is returned without a channel membership check.
- Open WebUI: Cross-user code-interpreter and tool execution via unvalidated Socket.IO event-caller session_id A user supplies another user's socket session id, so code and tool calls run in the victim's session and can reach admin-only server-side function execution.
- Open WebUI: Upload `metadata.knowledge_id` bypasses the knowledge-base write-access check (read-only users can add files to KB) An upload carrying a knowledge base ID is linked to that knowledge base before the write-permission check runs, letting read-only users add files.
- Open WebUI: Account enumeration via observable login timing discrepancy Login responses are slower for registered emails because password hashing runs only for existing accounts, revealing which addresses are registered.
- Open WebUI: Realtime endpoints accept Redis-revoked JWTs after signout/backchannel logout Realtime socket and terminal endpoints check only JWT signature and expiry, not the revocation store, so signed-out tokens can still connect.
- Open WebUI: ReDoS in skill-mention regexes causes whole-instance DoS on default config Skill-mention regexes backtrack quadratically on a long input with an unclosed tag, blocking the single worker's event loop.
- open-webui terminal proxy path traversal guard bypass via 9x encoded traversal Traversal sequences encoded nine times survive the proxy's limited decoding loop, so requests reach terminal server paths outside the permitted proxy prefix.
- Open WebUI: Unauthenticated WebSocket Access to Collaborative Document Handlers (ydoc:awareness:update, ydoc:document:leave) Socket.IO accepts unauthenticated connections, and two collaborative document handlers perform no session or membership checks.
- Open WebUI: Terminal proxy forwards a spoofable, integrity-unbound user identity to the upstream (X-User-Id header and ws_terminal session_id query injection) The terminal proxy forwards an unsigned user ID header and lets injected query characters in session_id override the user identity sent upstream.
- Open WebUI: Arena task endpoints can bypass underlying model access controls Arena task routes call the generator directly, so the access check on a restricted submodel is skipped once the wrapper is authorised.
- Open WebUI: Scheduled automations continue after pending-user deactivation and stored model ACL revocation Scheduled automations keep running for deactivated pending accounts because the scheduler skips owner status checks and model access logic mishandles that role.
- Open WebUI: `WEB_FETCH_FILTER_LIST` host allow/block filter bypassable via URL path and non-label-boundary matching The host allow and block filter matches full URL strings with endswith, so path suffixes and partial labels let authenticated users reach filtered hosts.
- Open WebUI: POST /api/v1/images/edit bypasses the global image-edit switch and the per-user image-generation permission An authenticated non-admin can call the image edit route, which skips the global edit switch and per-user permission, triggering billable provider requests.
- Open WebUI: Cross-channel message overwrite via chat completion API (single-model and multimodel message_ids) A channel-prefixed chat ID combined with a chosen message ID lets a user overwrite messages in channels they do not belong to.
- LiteLLM: Local file read via request-supplied OIDC file references A health test endpoint resolves request-supplied OIDC file references, letting a privileged caller make the server read arbitrary local files.
- LiteLLM: Arbitrary file write via path traversal in Skills archive extraction Skill ZIP entries containing traversal sequences are written outside the staging directory when the skill is processed.
- LiteLLM: Custom Code Guardrails production endpoints bypass code safety checks Production create and update routes for custom guardrail code skip the sandbox used by the test endpoint, so submitted Python runs inside the proxy.
- n8n: External Secrets Accessible via Workflow Expressions Outside Credentials Workflow node expressions resolve external secrets, letting project editors read their plaintext values without holding secrets permissions.
- n8n: Send Email Node Arbitrary File Read and SSRF via Nodemailer Content-Object Type Confusion Non-string values reaching a mail node's body fields are treated by the mail library as file paths or URLs, exposing local files from the host.
- n8n: Improper Authorization Allows Authenticated Users to Assign Workflows to Folders in Other Projects A crafted creation request links a new workflow to a folder in another project because the server does not verify folder ownership.
- n8n: Unauthenticated Endpoint Allows Cancellation of Any User's Active Test Webhook A route registered before auth middleware lets unauthenticated callers cancel active test webhooks by workflow ID.
- n8n: SSO Instance-Role Provisioning Allows Privilege Escalation to Instance Owner An SSO user whose identity provider sends an owner role claim is provisioned as instance owner, because the provisioning path skips the owner-role check.
- n8n: Snowflake Node executeQuery Operation Allows SQL Injection via Unparameterized Expression Interpolation Expression values from workflow data are concatenated into a raw Snowflake SQL string instead of bound as parameters, so attacker data can alter the query.
- n8n: Credential Authorization Bypass via Expression in HTTP Request Node `genericAuthType` A workflow editor supplies the HTTP node credential type through an expression, so the permission check misses and another user's credential loads.
- n8n: Prototype Pollution via Workflow Credentials Leads to Unauthenticated User and Project Enumeration A user with workflow create rights saves a workflow with prototype-polluting keys, so later unauthenticated requests are treated as privileged and user data is exposed.
- n8n: Privilege Escalation and Code Execution via Full Public API Key Scope Assignment to Token Exchange JWTs Tokens minted by the token exchange flow received every Public API scope regardless of the user's role, letting low-privileged callers perform administrator operations.
- n8n: Account Takeover via Unverified Email Claim in Token Exchange Embed Login A validly signed token is matched to a local account by email without checking the key's role ceiling or that the email is verified, enabling account takeover.
- n8n: Cross-Tenant Module-Cache Poisoning in the JS Task Runner Code-node users poison a shared JavaScript module cache, altering other tenants' later Code-node executions on the same task runner.
- n8n: Authenticated SSRF via Dynamic Node Parameters Endpoints Allows Internal Network Access Dynamic node parameter routes lack scope checks and accept absolute URLs that override the node base URL, letting users make the server call internal targets.
- n8n: Stored DOM XSS via Resource Locator `cachedResultUrl` A workflow's cachedResultUrl is passed to window.open without scheme checks, so a javascript URL runs script when a victim opens it.