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 Slack app_mention bypasses configured user/channel authorization The Slack mention handler lacks the user and channel allowlist checks present in the message handler, so unauthorised senders can trigger the agent.
- PraisonAI: Jobs webhook SSRF protection bypass via DNS rebinding A webhook hostname passes validation at submission but later resolves to an internal address when the server sends the callback, evading the earlier check through DNS rebinding.
- PraisonAI: Webhook signature verification skipped (fail-open) when secret unset, allowing forged inbound webhooks (WhatsApp & Linear bots) Webhook adapters skip HMAC signature checks whenever the secret is unset, so an attacker can post forged platform events that are dispatched as trusted.
- PraisonAI: Unauthenticated Event Injection via SSE `/publish` Endpoint An unauthenticated endpoint accepts arbitrary messages and broadcasts them to all SSE clients, and an info endpoint exposes server configuration even though an auth token is defined.
- praisonaiagents: SSRF guard validates literal IPs only and never resolves DNS The SSRF guard checks only literal IP forms of a URL host and never resolves DNS, so a hostname pointing to an internal address gets fetched.
- PraisonAI AgentTeam.launch exposes unauthenticated remote agent listing and invocation endpoints Agent-serving routes registered by the launch helper accept requests without checking credentials, so any network client can list and invoke agents.
- Langflow: Unauthenticated file upload leads to DoS (space exhaustion) and information leak The upload endpoint accepts unlimited unauthenticated data and returns the absolute server path, enabling disk exhaustion and path disclosure.
- Open WebUI: Redirect-Bypass SSRF in OAuth `_process_picture_url` (incomplete-fix sibling of CVE-2026-45401) The OAuth profile picture fetch validates only the initial URL and follows redirects, so a public URL can redirect the server to internal addresses.
- Open WebUI: Cross-origin postMessage confirmation bypass via action:submit The chat page message listener accepts prompt and submit messages from any origin, so an external page can submit prompts and trigger model requests without the confirmation dialog.
- Open WebUI Prompt history IDOR: unbound history_id allows cross-prompt read and deletion The prompt history endpoints check the parent prompt but not that the supplied history entry belongs to it, so another user's history can be read or deleted.
- Open WebUI: Any authenticated user can read other users' private notes via Socket.IO A Socket.IO join handler skips its ownership check when a note's document ID uses an underscore instead of a colon, and the storage layer maps both forms to the same document.
- Open WebUI: Authenticated users can target arbitrary configured Ollama backends via unguarded url_idx path parameter A signed-in user adds a raw backend index to Ollama proxy routes, so requests reach backends that their model permissions never covered.
- Open WebUI IDOR: Calendar event re-parenting allows writing events into another user's calendar The event update endpoint checks write access only on the source calendar, not the destination ID in the body, so events can be moved into others' calendars.
- Open WebUI: Sibling-Prefix Path Traversal via /cache/{path} A prefix-only directory check lacks a trailing separator, so sibling directories sharing the cache name prefix can be read by authenticated users.
- Open WebUI: SSRF Protection Bypass in Playwright Web Loader via HTTP Redirects The URL check runs only on the first address, so a page that redirects the headless browser to an internal host reaches blocked targets.
- Open WebUI: Path traversal / SSRF in terminal server proxy via encoded path traversal Double-encoded dot segments slip past the proxy's path sanitiser, letting an authenticated user reach files and endpoints outside the intended scope on the terminal server.
- Open WebUI: Stored XSS to Account Takeover via Model Profile Images Model profile image uploads accept SVG data URIs that earlier fixes missed, so embedded scripts run when the image URL is opened.
- Open WebUI: Stored XSS in Mermaid Markdown Preview A Markdown file with a Mermaid click directive carrying a javascript: URL is rendered into the DOM unsanitised, so script runs when the preview opens.
- Open WebUI: Forged chat-file link allows cross-user file read and deletion Message file IDs are stored without ownership checks, and a shared chat then grants access to those files, letting an attacker read or delete other users' files.
- Open WebUI: Cross-user file disclosure via /api/chat/completions image_url field An authenticated user passes another user's file ID as an image URL, and the server reads that file without an ownership check and sends its contents to the model.
- vLLM: GGUF dequantize kernel int truncation exposes uninitialized GPU memory in multi-tenant serving Integer truncation in GGUF dequantization leaves part of a newly allocated GPU buffer unwritten, exposing residual memory from earlier requests.
- vLLM: incomplete CVE-2026-22778 fix leaks PIL repr addresses via Anthropic router Malformed image bytes make the Anthropic-compatible router echo an unsanitised exception containing a heap address, because a later endpoint skipped the earlier fix.
- n8n: Wrong OAuth Scope on Evaluation Test Runs Endpoints Evaluation test-run endpoints check the read scope instead of the execute scope, letting viewers start, cancel and delete runs on workflows.
- Langflow: Path Traversal in Knowledge Bases API via Creation Endpoint The knowledge base name in the create endpoint is joined into a filesystem path without sanitisation, letting an authenticated user create directories and write files outside their area.
- Langflow: IDOR/BOLA in Monitor API — Missing Ownership Enforcement on 7 Endpoints An authenticated user reads, modifies or deletes other users' monitor records by supplying their resource or flow IDs, because several endpoints skip ownership checks.
- Langflow: Unauthenticated Shareable Playground arbitrary local or S3 file read The unauthenticated public flow endpoint accepts a files list whose paths are read without restriction, so anyone with the link can read local or S3 files.
- Langflow: Unauthenticated RCE in Shareable Playgrounds A public flow endpoint accepts replacement Python code in node definitions and runs it without authentication.
- LiteLLM: Authentication Bypass via Host Header Injection A crafted Host header makes the proxy's auth gate check a different route from the one served, skipping authentication on management endpoints.
- Natural Language Toolkit (NLTK): URL-Encoded Path Traversal in nltk.data.load() Allows Arbitrary Local File Read The traversal check runs on the raw resource name before percent-decoding, so encoded separators and dot segments reach the filesystem and read arbitrary files.
- vLLM: OpenAI auth bypass A crafted Host header changes the path the auth middleware parses, so /v1 requests skip the API key check while still being routed.
- n8n: Cross-Tenant Credential Takeover via Dynamic Credentials EE Endpoints Dynamic credential endpoints skip ownership checks, so any logged-in user can read or overwrite another tenant's credential tokens.
- n8n: Prototype Pollution enables confused-deputy execution via public webhooks Injected prototype properties in public webhook data flow into workflow fields, steering actions that run with the owner's credentials.
- n8n: Stored XSS in Chat Trigger Node A user with workflow edit rights sets a malicious webhook ID that injects script into the chat page, which then runs for anyone who opens it.
- n8n: Git Node Clone and Push Operations Bypass File Sandbox The Git node accepts local filesystem paths as clone or push sources, letting workflows copy repositories that the file sandbox would block.
- n8n: Merge Node SQL Mode Prototype Pollution A workflow author pollutes a cached sandbox prototype used by the Merge node SQL mode, and the polluted state carries over to other users' later executions, exposing their data.
- n8n: Python sandbox escape Authenticated users with workflow edit rights can break out of the Python code node sandbox and run commands in the task runner container.
- n8n: SQL Injection in Postgres v1/TimesclaeDB Nodes Parameters for the Postgres and TimescaleDB workflow nodes are concatenated into SQL, letting a workflow author run arbitrary queries on the connected database.
- n8n: Public API Execution Retry Authorization Bypass The public API retry endpoint checks workflow read permission instead of execute permission, so read-only sharing users can re-run executions.
- n8n: Reflected XSS via Facebook, WhatsApp, and Microsoft Teams Trigger Webhook Verification Endpoints A webhook verification endpoint echoes a query parameter into the response without escaping or CSP, so a crafted link runs script in the app origin.
- n8n: Wrong OAuth Scope On Evaluations Test Run Creation Endpoint The test-run endpoint checks only read permission, so users who can view a workflow can start an execution they are not allowed to run.
- n8n: NoSQL Injection in MongoDB Node Find And Replace Operation An unvalidated filter value lets a workflow editor match and overwrite MongoDB documents beyond the intended set.
- n8n: Denial of Service via ZIP decompression in webhook workflow A small compressed archive sent to a public webhook expands without limits and exhausts the process memory.
- n8n: Missing Token Validation on Microsoft Agent 365 Trigger and Stripe Nodes Two webhook trigger nodes accept inbound requests without verifying a token, so anyone who knows the URL can start workflows with forged data.
- n8n: Python Code Node AST Validator Bypass An authenticated user can bypass the AST check on Python code nodes to reach the task executor namespace and read environment variables.
- n8n: Credential Exfiltration via Permission Bypass Some public API routes resolve credentials without full ownership checks, so an editor can reference and use credentials belonging to other users.
- n8n: SecurityScorecard Node Leaks API Token to User-Controlled Host A workflow author points the SecurityScorecard report download at an attacker host, and the node sends the stored API token there despite the configured domain limits.
- n8n: Same-Origin XSS in Respond to Webhook Node A binary response path serves attacker-chosen content types without the sandbox header, so a public webhook runs scripts in the application origin.
- n8n: Microsoft SQL Node Prototype Pollution A crafted table parameter in the Microsoft SQL node pollutes Object.prototype process-wide, breaking the server until restart.
- Keras archive extraction utilities allow path traversal and arbitrary file writes Archive filters check member paths against the working directory rather than the extraction target, so traversal entries escape when the process runs from root.
- songquanpeng one-api has an issue that results in business logic errors A crafted request to the redemption endpoint triggers business logic errors in redeem handling because input is not validated properly.