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
- VM2 Sandbox Breakout Through __lookupGetter__ A getter lookup inside vm2 returns host-side objects, letting sandboxed code reach the host Function constructor and run commands outside the sandbox.
- Claude SDK for TypeScript has Insecure Default File Permissions in Local Filesystem Memory Tool The local filesystem memory tool creates agent memory files with permissive default modes, so other local users can read them and in containers may modify them.
- n8n has a Python Task Runner Sandbox Escape Vulnerability An authenticated workflow author can break out of the Python task runner sandbox used by a Code node and run commands on the runner container.
- n8n has Public API Variables IDOR that Allows Cross-Project Secret Disclosure An API key scoped to variable listing passes an arbitrary project id, and the handler returns variables without checking project membership.
- n8n Vulnerable to Hijacking of Unauthenticated Chat Execution The chat WebSocket does not check authorisation, so an unauthenticated client who knows a waiting execution ID can attach to it and submit input.
- n8n has SQL Injection in Snowflake and MySQL Nodes Table, column and update-key identifiers are interpolated into SQL in the Snowflake and legacy MySQL nodes, a gap the earlier fix did not cover.
- n8n has XML Node Prototype Pollution that to RCE XML node parsing lets an authenticated workflow author pollute global object prototypes, which can chain with other nodes into remote code execution.
- n8n has SQL Injection in SeaTable Node Unescaped user input concatenated into SeaTable row search queries lets an attacker alter the query and retrieve rows the workflow meant to filter out.
- n8n has Prototype Pollution in XML Webhook Body Parser that Leads to RCE Crafted XML webhook bodies pollute JavaScript object prototypes in the parser, and the pollution is chained with the Git node to run commands on the host.
- n8n's Credential Authorization Bypass in dynamic-node-parameters Allows Foreign API Key Replay A user with shared workflow access supplies a foreign credential ID to dynamic parameter endpoints, making the server send another user's API key to an attacker-chosen URL.
- n8n has SQL Injection in Oracle Database Node via Limit Field Input flowing through an expression into the Oracle node's Limit field is concatenated into the SQL query, letting webhook-triggered attackers run arbitrary queries.
- vLLM makes Use of Uninitialized Resource A helper in vLLM's KV cache interface reads an uninitialised resource, reachable remotely though hard to exploit.
- Ollama is Vulnerable to Path Traversal A crafted digest value in the model transfer handler is not sanitised before path construction, letting requests reach files outside the intended directory.
- New API: Stripe Webhook Signature Bypass via Empty Secret Enables Unlimited Quota Fraud An empty Stripe webhook secret makes signatures trivially computable, and orders from other gateways can be fulfilled, crediting unpaid quota.
- LiteLLM has SQL Injection in Proxy API key verification An Authorization header value is concatenated into a key-verification query on error paths, so unauthenticated requests can inject SQL against the proxy database.
- LiteLLM: Server-Side Template Injection in /prompts/test endpoint A user-supplied prompt template is rendered without a sandbox, so crafted template expressions can execute code inside the proxy process.
- H2O-3 is Vulnerable to Code Injection A parameter blocklist meant for MySQL JDBC is avoided by switching to the PostgreSQL driver, whose socket factory options run attacker-supplied code.
- AgentScope vulnerable to Server-Side Request Forgery The internal helper that downloads bytes from a web URL accepts attacker-chosen addresses, letting requests reach internal services.
- AgentScope vulnerable to Server-Side Request Forgery The audio block processor fetches a caller-supplied URL without restriction, so it can be made to request internal resources.
- Langflow: DoS Through Lack of File Size Restriction via Deprecated Unauthenticated File Upload API The deprecated unauthenticated upload endpoint applies no size limit, so remote clients can fill disk or memory and cause denial of service.
- Langflow: Cleartext Storage of Authentication Settings in Project Creation Endpoint The project creation endpoint stores authentication settings in cleartext on disk instead of encrypting them.
- PraisonAI: SQL Injection via unvalidated `table_prefix` in 9 conversation store backends (incomplete fix for CVE-2026-40315) The table prefix is interpolated into SQL in nine store backends that lack the validation added to the SQLite store, so crafted prefixes inject queries.
- Claude Code: Insecure System-Wide Configuration Loading Enables Local Privilege Escalation on Windows A low-privileged Windows user creates the ClaudeCode folder under ProgramData and places a settings file that other users' sessions then load automatically.
- Flowise: Mass Assignment in DocumentStore Create Endpoint Leads to Cross-Workspace Object Takeover (IDOR) The create endpoint accepts a client-supplied primary key and saves the whole body, so it overwrites existing document stores owned by other workspaces.
- Flowise: Unauthenticated TTS endpoint accepts arbitrary credential IDs — enables API credit abuse via stored credentials The TTS generate route skips authentication and accepts any credential ID in the body, letting callers spend stored provider API keys belonging to others.
- langchain-openai: Image token counting SSRF protection can be bypassed via DNS rebinding Image URLs are validated for private addresses and then fetched separately, so DNS rebinding lets the fetch reach internal hosts despite the check.
- LangChain Text Splitters: HTMLHeaderTextSplitter.split_text_from_url SSRF Redirect Bypass URL validation runs only on the first request, and redirects are then followed without revalidation, so an attacker can reach internal or metadata endpoints.
- Flowise: SSRF Protection Bypass (TOCTOU & Default Insecure) DNS rebinding exploits the gap between a validation lookup and the connection lookup, and the default empty deny list lets requests reach localhost.
- Flowise Execute Flow function has an SSRF vulnerability The Execute Flow node's base URL is not checked by the deny-list validator, so the server issues requests to internal addresses.
- Flowise: Code Injection in CSVAgent leads to Authenticated RCE Unsanitised custom Pandas read code is interpolated into Python run by pyodide in the CSV agent, so a crafted chatflow executes OS commands.
- Flowise: Parameter Override Bypass Remote Command Execution A substring check meant for file inputs accepts the keyword anywhere in a request, letting unauthenticated users override protected config such as NODE_OPTIONS and run commands.
- Flowise: SSRF Protection Bypass via Direct node-fetch / axios Usage (Patch Enforcement Failure) Several tools call raw HTTP clients directly instead of the centralised SSRF wrapper, so requests to internal addresses still succeed.
- Flowise: Path Traversal in Vector Store basePath The basePath parameter for vector store saves is used directly in filesystem writes, so authenticated users can write outside the intended directory.
- Flowise: SSRF Protection Bypass via Unprotected Built-in HTTP Modules in Custom Function Sandbox Custom function code can load built-in http, https or net modules that bypass the SSRF deny list, reaching internal services such as cloud metadata endpoints.
- Flowise: Weak Default Express Session Secret A hardcoded fallback session secret lets anyone forge signed session cookies and impersonate other users.
- Flowise: Improper Mass Assignment in Account Registration Enables Unauthorized Organization Association The registration endpoint maps unfiltered client JSON onto internal models, letting attackers set organization, workspace and role fields during signup.
- Flowise: Unauthenticated OAuth 2.0 Access Token Disclosure via Public Chatflow in Flowise A public endpoint exposes OAuth credential identifiers from chatflow data, and an unauthenticated refresh endpoint then issues access tokens.
- Flowise: Unauthenticated Information Disclosure of OAuth Secrets (Cleartext) via GET Request An unauthenticated endpoint returns an organisation's SSO provider configuration, including client secrets in plain JSON, to any caller.
- Flowise: Weak Default JWT Secrets Hardcoded fallback JWT secrets let anyone who knows the defaults forge tokens for any account.
- Flowise: resetPassword Authentication Bypass Vulnerability The account service's password reset logic can be reached without authentication, letting remote attackers change other users' passwords.
- Flowise: Weak Default Token Hash Secret A hardcoded fallback secret is used to encrypt identifiers in tokens, so anyone who knows the default can forge tokens.
- Flowise: File Upload Validation Bypass in createAttachment Adding a JavaScript MIME type to chatflow upload settings lets an attacker store a Node web shell that then runs on the server.
- Flowise: Password Reset Link Sent Over Unsecured HTTP Password reset emails contain plain HTTP links, so a network attacker on the path can capture the token and take over accounts.
- Google Agent Development Kit (ADK) has a Code Injection and Missing Authentication vulnerability An unauthenticated caller can reach a route in the ADK server that runs attacker-supplied code on the host.
- MetaGPT affected by server-side request forgery in metagpt/utils/common.py The decode_image helper fetches attacker-supplied URLs without restriction, letting callers make the server request internal or arbitrary addresses.
- MetaGPT has an eval injection via a cross-site request forgery attack A cross-site request forgery against the Mineflayer HTTP API triggers evaluation of attacker-supplied code in the agent's environment.
- parisneo/lollms vulnerable to stored XSS in the social feature Social feed post content is stored without sanitisation and rendered to viewers, so injected script runs in their browsers.
- PraisonAI has Unrestricted Upload Size in WSGI Recipe Registry Server that Enables Memory Exhaustion DoS The registry server reads request bodies fully into memory based on the client's Content-Length with no upper limit, so large uploads exhaust memory.
- PraisonAI Vulnerable to Server-Side Request Forgery via Unvalidated webhook_url in Jobs API The jobs API posts completion results to any caller-supplied webhook URL without validation, letting the server reach internal and cloud metadata endpoints.
- PraisonAI vulnerable to arbitrary file write via path traversal in `praisonai recipe unpack` A malicious .praison archive with ../ entries is extracted with raw tar extraction, letting it write files outside the target directory.