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
- PraisonAI Vulnerable to Decompression Bomb DoS via Recipe Bundle Extraction Without Size Limits Recipe archives are extracted with path checks but no limits on member size or count, so a small, highly compressible bundle can fill the victim's disk.
- PraisonAI Vulnerable to Argument Injection into Cloud Run Environment Variables via Unsanitized Comma in gcloud --set-env-vars Commas in model or key values are not escaped when building the gcloud env-var argument, so attackers can inject extra environment variables into the deployed service.
- PraisonAI: SQLiteConversationStore didn't validate table_prefix when constructing SQL queries The table prefix from configuration is concatenated into SQL identifiers without validation, letting injected fragments read internal tables.
- PraisonAI: Cross-Origin Agent Execution via Hardcoded Wildcard CORS and Missing Authentication on AGUI Endpoint The agent endpoint has no auth and sends a wildcard CORS header, so any website can run agents and read the responses.
- LangSmith Client SDKs has Prototype Pollution in langsmith-sdk via Incomplete `__proto__` Guard in Internal lodash `set()` An incomplete __proto__ guard in a vendored lodash set lets keys using constructor.prototype pollute Object.prototype in the Node.js process.
- FoundationAgents MetaGPT vulnerable to OS Command Injection in metagpt/utils/common.py The MIME type helper builds an operating system command from untrusted input, allowing injected shell commands to execute.
- LangChain has incomplete f-string validation in prompt templates Template validation misses attribute access and nested format-spec fields, so attacker-supplied f-string templates can evaluate expressions that expose object attributes during formatting.
- LiteLLM: Password hash exposure and pass-the-hash authentication bypass User endpoints return stored password hashes to any authenticated user, and login accepts that hash in place of the password, allowing account takeover.
- parisneo/lollms has an insufficient session expiration vulnerability Existing session tokens remain valid after a password reset and last about 31 days, so a stolen token keeps working.
- PraisonAI Vulnerable to Remote Code Execution via YAML Deserialization in Agent Definition Loading A crafted agent YAML file uploaded through the API is parsed with an unsafe loader that evaluates embedded JavaScript functions.
- PraisonAI Has Unauthenticated SSE Event Stream that Exposes All Agent Activity in A2U Server Agent event streaming endpoints have no authentication, so anyone can open a subscription and receive live responses, reasoning and tool call arguments.
- PraisonAI has Memory State Leakage and Path Traversal in MultiAgent Context Handling The monitor builds file paths from agent ids without removing traversal sequences, and agents sharing an id also share one ledger, leaking context between them.
- LoLLMs is vulnerable to Improper Access Control through weak secret key A guessable JWT signing key can be brute-forced offline, and the recovered key lets an attacker forge admin tokens to reach restricted endpoints.
- MLflow is vulnerable to an authorization bypass affecting the AJAX endpoint The MLflow AJAX endpoint for model artifacts skips the experiment permission check, so users without access can download artifacts directly.
- MLflow is vulnerable to Stored Cross-Site Scripting (XSS) caused by unsafe parsing of YAML-based MLmodel artifacts in its web interface Unsafe YAML parsing of MLmodel metadata lets an uploaded artifact inject script that runs in another user's browser when viewed in the UI.
- PraisonAI Has Arbitrary File Write (Zip Slip) in Templates Extraction Template archive entries with parent-directory segments are extracted without path checks, letting a malicious template overwrite files outside its folder.
- PraisonAI recipe registry pull path traversal writes files outside the chosen output directory A recipe bundle whose tar archive contains parent-directory entries makes the client extract files outside the chosen output folder when pulled.
- PraisonAI recipe registry publish path traversal allows out-of-root file write The registry writes uploaded bundles using a manifest-supplied name and version before validating them, so traversal sequences write files outside the registry root.
- BentoML: Command Injection in cloud deployment setup script Unquoted system_packages values from bentofile.yaml are placed in a setup script run on cloud build infrastructure, a path an earlier quoting fix missed.
- Kedro: Path Traversal in versioned dataset loading via unsanitized version string A version string containing ../ sequences is joined into the versioned dataset path without validation, letting loads read files outside the intended directory.
- Kedro has Arbitrary Code Execution via Malicious Logging Configuration Setting the logging config path to a file containing a special factory key makes dictConfig call an arbitrary callable at application startup.
- LiteLLM: Privilege escalation via unrestricted proxy configuration endpoint The config update route lacks an admin role check, so any logged-in user can change proxy settings, register code handlers and overwrite admin credentials.
- LiteLLM: Authentication bypass via OIDC userinfo cache key collision The OIDC userinfo cache keys on the first 20 characters of the JWT, so a forged token sharing that prefix gets the cached identity of a legitimate user.
- mlflow: FastAPI job endpoints under `/ajax-api/3.0/jobs/*` are not protected by authentication or authorization Job submission and management endpoints accept requests without credentials when basic-auth is enabled, letting any network client create and cancel jobs.
- vLLM: Server-Side Request Forgery (SSRF) in `download_bytes_from_url ` The batch runner fetches any URL supplied in batch input JSON without validation, letting callers reach internal hosts and metadata endpoints.
- ONNX: TOCTOU arbitrary file read/write in save_external_dat The external data writer checks whether a path is a file and then opens it separately, letting an attacker swap in a symlink to overwrite arbitrary files.
- Open WebUI has Broken Access Control in Tool Valves A regular member reads tool valve configuration, including third-party API keys, through an endpoint that does not restrict read access to admin-only values.
- PraisonAI Has Second-Order SQL Injection in `get_all_user_threads` A thread ID stored with a crafted SQL fragment is later concatenated unescaped into a thread-listing query, executing injected SQL against the database.
- PraisonAI Has Missing Authentication in WebSocket Gateway The gateway accepts WebSocket connections and lists agents without authentication, letting any network client send messages to agents and their tools.
- PraisonAI: SSRF via Unvalidated api_base in passthrough() Fallback The passthrough fallback joins a caller-supplied base URL to the endpoint and sends the request with no host validation, reaching internal services.
- Mlflow: Command Injection when serving models with enable_mlserver=True The model URI is interpolated unescaped into a bash -c command, so shell command substitutions inside it run when the model is served.
- MLFlow path traversal vulnerability A tar archive whose member paths contain traversal sequences is extracted without validation, so files can be written outside the target directory.
- LangChain Core has Path Traversal vulnerabilites in legacy `load_prompt` functions Prompt config loaders read template and example files from paths in config dicts without blocking traversal or absolute paths.
- Langflow: Authenticated Users Can Read, Modify, and Delete Any Flow via Missing Ownership Check Flow read, update and delete routes query by flow ID alone with no ownership filter, letting any authenticated user act on other users' flows.
- MLFlow allows Tracing + Assessments Access Trace and assessment endpoints skip permission checks when basic auth is enabled, so any authenticated user can read traces and create assessments.
- Open WebUI has unauthorized deletion of knowledge files The knowledge file removal endpoint checks write access to the knowledge base but not that the file belongs to it, so files can be deleted from other knowledge bases.
- Open WebUI's process_files_batch() endpoint missing ownership check, allows unauthorized file overwrite A batch file-processing route checks only that the caller is logged in, letting any user overwrite stored file contents by UUID.
- Open WebUI vulnerable to Path Traversal in `POST /api/v1/audio/transcriptions` A crafted multipart filename makes the extension include slashes, so the write fails and the error response leaks the server's absolute data directory path.
- Open WebUI's Insecure Direct Object Reference (IDOR) allows access to other users' memories Any logged-in user can query other users' collection names, including memory collections, through the retrieval endpoint because ownership is never checked.
- n8n has XSS in its Credential Management Flow A crafted OAuth2 credential with a javascript: URL in the authorization URL field runs script in the browser of a victim who clicks the authorize button.
- n8n has XSS in Chat Trigger Node through Custom CSS Custom CSS in the Chat Trigger node is stored after a misconfigured sanitize-html pass, allowing script to run on the public chat page.
- n8n has a Stored XSS Vulnerability in its Form Trigger A flaw in CSS sanitisation of Form Trigger nodes lets a stored payload run in every visitor's browser of the published form.
- n8n: Authenticated XSS and Open Redirect via Form Node A workflow editor can store HTML in a form description or exploit a loose iframe sandbox, so visitors get script execution or redirects.
- BentoML has Dockerfile Command Injection via system_packages in bentofile.yaml The system_packages values in bentofile.yaml are placed unquoted into Dockerfile RUN commands, so shell metacharacters in a package name run commands during the build.
- n8n has SQL Injection in Data Table Node via orderByColumn Expression Expressions in the Data Table Get node's orderByColumn field are concatenated into SQL, letting authenticated workflow editors inject queries.
- n8n: Prototype Pollution in XML and GSuiteAdmin node parameters lead to RCE A crafted node parameter in a workflow pollutes Object.prototype, and the pollution can be escalated to code execution on the server.
- n8n Vulnerable to XSS via Binary Data Inline HTML Rendering Binary data without a filename is served inline on the main origin without download or CSP headers, so crafted HTML runs script in a victim's session.
- n8n Vulnerable to LDAP Filter Injection in LDAP Node Unescaped LDAP metacharacters in user input reach the search filter of the n8n LDAP node, letting attackers alter queries and bypass workflow checks.
- n8n's Source Control SSH Configuration Uses StrictHostKeyChecking=no Source control SSH disables host key verification, which lets a network attacker impersonate the Git server and intercept repository data.
- n8n has Multiple Remote Code Execution Vulnerabilities in Merge Node AlaSQL SQL Mode The AlaSQL sandbox behind the Merge node's SQL mode fails to block certain statements, letting an authenticated workflow editor read host files and run code.