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: Shared-Workflow Editor Can Exfiltrate Credentials via Inline Sub-Workflow JSON Credential checks inspect only top-level node credentials, so an editor can reference other credentials inside a sub-workflow's inline JSON and use them.
- n8n: Race Condition in Git Clone Node Allows Authenticated Users to Achieve Remote Code Execution A race between validating the clone path and running it lets a user swap in a symlink, planting a repository where the server loads custom nodes as code.
- n8n: Git Node fetch/pull/pushTags Operations Bypass Sandbox Path Restriction A workflow author can point an allowlisted git remote at a local repository outside the sandbox, fetch it into the workspace, and read its files.
- n8n: GraphQL Node Bypasses "Allowed HTTP Request Domains" Restriction The GraphQL node skips the allowed-domains restriction for HTTP credentials, so a workflow author can send stored credentials to an attacker-controlled host.
- n8n: Expression sandbox escape via arrow-function bodies enabling command execution Crafted arrow-function bodies inside n8n expressions escape the expression sandbox and reach host system commands.
- n8n: MySQL v1 Node executeQuery Operation Allows SQL Injection via Unparameterized Expression Interpolation Expression values from webhook input are concatenated into a raw MySQL query in the legacy node, letting attacker-supplied SQL run.
- n8n: Prototype Pollution via VM Expression Engine Sandbox Escape Leads to Denial of Service An authenticated user's expression escapes the VM engine through array access, reaches a host built-in and pollutes its prototype, causing denial of service in the main process.
- n8n: External Secrets Permission Bypass via Expression Parser Mismatch Expression parsing misses some secret-reference forms, so a user can embed secrets they lack permission to read into credentials.
- n8n: PostgresTrigger Node SQL Injection Allows Authenticated Users to Execute Arbitrary SQL on Connected PostgreSQL Instances Unescaped channel, function and trigger names let an authenticated user run arbitrary SQL against the connected database.
- n8n: Cross-Issuer Token Exchange Account Binding via Subject-Only Identity Resolution Token exchange maps external identities to local accounts by subject alone, so a token from another trusted issuer with a matching subject logs in as the victim.
- n8n: Legacy Expression Evaluator Sanitizer Bypass Leads to Authenticated Code Execution An authenticated workflow editor can get past the legacy expression evaluator's member-access sanitizer and run code as the server process.
- n8n: Shared Credential Header Leak via HTTP Request Pagination Expression A user with use-only workflow access reads a shared HTTP Header Auth secret from pagination request state, then copies it into item data and sends it onward.
- n8n: Authenticated code execution in the n8n Git node A workflow user supplies a crafted local Git repository whose hooks run when the Git node operates on it, executing commands as the server process.
- n8n: Authenticated Users Can Exhaust Temporary Disk Storage via Data-Table File Uploads Repeated data-table uploads each pass the per-request quota because earlier temporary files are not counted, filling the disk until periodic cleanup runs.
- n8n: Edit Image Node Format Injection Allows Arbitrary File Write The Edit Image node passes its output format unvalidated to the image library, letting a crafted value write bytes outside the node's working directory.
- n8n: Prototype Pollution via Dot-Notation Field Names Leads To Instance-Wide Denial of Service A dot-notation field name in the Set node pollutes a shared global used in authentication, causing all authenticated requests to fail.
- vLLM: Speech-to-text upload size limit is enforced after full UploadFile read The speech-to-text upload handlers read the whole file before the documented size limit is checked, so oversized uploads consume memory first.
- Keras: tar extraction permits symlink-based path traversal Symlink entries in a crafted tar archive skip the directory containment check, so they can point outside the extraction folder and enable reads and overwrites.
- New API is vulnerable to CSRF through user email binding Account binding actions use GET requests, so a cross-site navigation can make a logged-in user's browser bind attacker-controlled email or OAuth identities.
- New API: SSRF Protection Bypass via Unresolved Hostname in Notification URLs Notification URL hostnames are not resolved or checked against private IPs by default, so authenticated users can reach internal addresses.
- LocalAI POST /models/apply permits unauthenticated server-side request forgery through gallery URLs An unauthenticated model-apply endpoint fetches attacker-supplied gallery URLs, letting callers make the server request internal addresses and see partial responses in errors.
- Open WebUI allows limited stored XSS vila uploaded html file A low-privileged user uploads an HTML file containing script that runs in another user's browser when the file is opened.
- Open WebUI has Blind Server Side Request Forgery in its Image Edit Functionality The image edit endpoint fetches any user-supplied HTTP or HTTPS URL without restriction, letting the server reach internal addresses that can be probed by response differences.
- MLflow: trace API endpoints lack proper authorization validators Trace API routes skip authorization validators, so any authenticated user can read or change traces in experiments they cannot access.
- Gradio FileExplorer preprocess path traversal allows files outside root_dir to reach callbacks Crafted path segments make the FileExplorer preprocess step discard the root directory prefix, letting unauthenticated callers read files outside it.
- Langflow: Code Injection Vulnerability in Code Validation Endpoint The code validation endpoint lets authenticated users inject Python that the server executes, allowing OS commands and reading of sensitive files.
- pydantic-ai: SSRF blocklist bypass via IPv4-compatible, SIIT/IVI, and local NAT64 IPv6 addresses (incomplete fix of CVE-2026-46678) IPv6 transition address forms evade the private-address blocklist during URL downloads, letting the library reach cloud metadata endpoints and expose credentials.
- @anthropic-ai/claude-code has an Insecure Temporary File in /copy Command that Enables Response Disclosure and Symlink-Based File Write The copy command writes responses to a predictable, world-readable temp path, so local users can read them or redirect the write through a planted symlink.
- Feast vulnerable to unsafe deserialization RCE in registry server via UDF proto The registry server runs dill.loads on a base64 UDF body before any authorisation check, so a crafted object with a malicious __reduce__ executes OS commands.
- Chainlit contains a session hijacking vulnerability Session restoration only checks that a session ID exists and not who owns it, so an unauthenticated caller can adopt a victim's session and tool permissions.
- ComfyUI-Manager has an Unprotected Alternate Channel (CWE-420) Configuration files in a web-reachable user directory can be read and modified by remote clients without any access control.
- LiteLLM: Admin Key Handler Has Improper Authorization A remote caller could reach a LiteLLM key management handler that fails to verify authorization, enabling administrative key operations.
- LiteLLM: SSO Debug Flow Has Improper Authentication The SSO debug endpoint lacks an authentication check, so remote callers can reach it without credentials.
- BerriAI litellm: UI User Enumeration leads to System-Wide Information Disclosure A user listing endpoint returns all accounts without proper authorization, and an earlier incomplete fix left it exposed.
- LiteLLM: PROXY_ADMIN database API Key Generator Has Insufficient Session Expiration Sessions created through the proxy admin API key flow do not expire properly, so a stale session can remain usable.
- LiteLLM: M2M JWT Handler Has Improper Authorization A flaw in the machine-to-machine JWT handler causes improper authorization, letting remote attackers gain access they should not have.
- BerriAI litellm: SSO Login Does Not Invalidate Previous UI Session Tokens Completing a new SSO login does not invalidate session tokens issued earlier for the UI, so old sessions stay usable.
- Langflow: Logout button does not clear session The logout endpoint fails to delete auth cookies with their original attributes, so the previous session stays valid in the browser.
- Langflow: IDOR Vulnerability in `/api/v1/responses` Endpoint Allows Authenticated Attackers to Access Another User's Flow Flows fetched by UUID are not checked against the requesting user, so any authenticated user can run another user's flow.
- Langflow: Unauthenticated DoS through multipart form boundary file upload The upload endpoint parses multipart data before authentication and does not limit extremely long boundaries, tying up the server for all users.
- LangSmith SDK TracingMiddleware: Arbitrary server-side file read Header-supplied fields inject file attachments, and a broken type check fails to stop the tracing thread reading any server file the attacker names.
- PraisonAI recipe serve Typer command bypasses the non-localhost authentication guard The deprecated Typer command starts the recipe server on any interface without attaching auth middleware, skipping the non-localhost guard.
- PraisonAI: AgentOS remains unauthenticated after incomplete fix version and allows remote agent invocation Agent API routes accept requests without authentication, so any caller can invoke deployed agents, including tool-enabled ones, through the chat endpoint.
- PraisonAI: PRAISONAI_CALL_AUTH=disabled environment variable unconditionally disables authentication Setting an environment variable to disabled makes the token verification function return early, so agent invocation requests need no credentials.
- PraisonAI GitHub template cache path traversal allows outside-cache file write and directory deletion Crafted GitHub template ref values escape the cache directory, letting the loader write files and delete arbitrary outside directories.
- PraisonAI LinearBot processes unsigned webhooks when LINEAR_WEBHOOK_SECRET is missing The webhook handler skips signature verification when the secret is unset, so forged events reach the agent session.
- PraisonAI: Jobs API exposes agent-execution endpoints with no authentication Job submission endpoints for running agents have no authentication, so any network caller can queue agent runs that execute attacker-supplied prompts and YAML.
- praisonai: recipe serve auth middleware silently disables itself when no secret is set The recipe server's auth middleware allows every request when no secret is configured, so enabling auth without setting a key leaves the API open.
- PraisonAI A2U incomplete authentication fix leaves current serve command unauthenticated by default The A2U serve command exposes event streams without authentication, because the auth check is skipped whenever the token variable is unset.
- PraisonAI: Unauthenticated Local File Inclusion via agent_file path in PraisonAI Jobs API An unauthenticated caller submits an absolute filesystem path to the Jobs API, and the server reads that file without any path validation.