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
- LlamaIndex Improper Handling of Exceptional Conditions vulnerability A wrong-typed input makes a worker thread exit early, and the streaming handler then waits forever for a response that never arrives.
- LlamaIndex Uncontrolled Resource Consumption vulnerability A URL pointing at a site root makes the web reader recurse endlessly through article links, exhausting memory and crashing the application.
- MLflow Uncontrolled Resource Consumption vulnerability Large batches of GraphQL queries repeatedly requesting all runs of an experiment tie up every MLflow worker, blocking other requests.
- MLflow has Weak Password Requirements Admins can create user accounts without setting a password, leaving those accounts open to anyone who can guess the username.
- MLflow has a Local File Read/Path Traversal in dbfs The dbfs URL handling checks only the path component, so traversal sequences in query or parameter parts escape the mount and read local files.
- MLflow Cross-Site Request Forgery (CSRF) vulnerability The signup endpoint accepts cross-site form submissions without anti-CSRF checks, so a third-party page can make a victim's browser create accounts.
- MLflow Uncontrolled Resource Consumption vulnerability Experiment names and artifact locations have no length limit, so oversized values make the MLflow UI hang and create resource exhaustion.
- Open WebUI Unauthenticated Multipart Boundary Denial of Service (DoS) Vulnerability Appending characters to a multipart boundary forces unauthenticated upload parsers to scan each byte, exhausting CPU and memory on several endpoints.
- Open WebUI denial of service through endpoint for converting markdown An unauthenticated markdown-to-HTML endpoint can be sent payloads that take excessive time to convert, tying up the server.
- Open WebUI Unauthenticated Multipart Boundary Denial of Service (DoS) Vulnerability in api/chat/file A malformed multipart boundary with many trailing characters makes the upload handler process each character endlessly, hanging the server for all users.
- Open WebUI Cross-Site Request Forgery (CSRF) Vulnerability Missing CSRF protection on a cookie-authenticated route lets a malicious page edit pipeline Python code, which then runs with the victim's privileges.
- Open WebUI lacks authentication for the `api/v1/utils/pdf` endpoint The PDF generation endpoint accepts requests without authentication, so anyone can submit oversized payloads that exhaust server resources.
- Open WebUI Has Improper Access Control Leading to Arbitrary Prompt Read The prompts API does not check for administrator rights, so any user can list other users' prompts and fetch them by command ID.
- Open WebUI Uncontrolled Resource Consumption vulnerability An unauthenticated endpoint accepts very large content for formatting, so an anonymous client can exhaust server resources until it stops responding.
- Open WebUI allows Remote Code Execution via Arbitrary File Upload to /audio/api/v1/transcriptions The upload endpoint trusts the client content type and filename, letting paths escape the upload directory and overwrite system files.
- Open WebUI Uncontrolled Resource Consumption vulnerability The sign-in form accepts unbounded email and password values, so oversized requests exhaust server CPU, memory and disk without authentication.
- Open WebUI stored cross-site scripting (XSS) vulnerability The model description set through the add model endpoint is stored with insufficient sanitising and later runs as script when rendered in chat.
- Open WebUI Allows Viewing of Admin Details The admin details endpoint does not check whether the caller is an administrator, so any caller can read the first admin's account details.
- Open WebUI Vulnerable to Cross-Site Scripting (XSS) via Chat File Upload An uploaded chat file carries script content that runs in a victim's browser when the file is opened through a link or shared chat.
- Open WebUI Allows Arbitrary File Reading and Deletion Any logged-in user can list, read and delete other users' uploaded files because the file endpoints never check for administrator rights.
- Open WebUI Vulnerable to Cross-Site Request Forgery (CSRF) Destructive reset and delete actions run on GET requests, so a malicious page or top-level link can trigger them in a logged-in user's browser.
- Open WebUI has vulnerable dependency on starlette via fastapi A bundled older starlette version, pulled in through fastapi, can be driven to exhaust memory by sending crafted requests.
- Open WebUI Uncontrolled Resource Consumption vulnerability An oversized name submitted at signup makes the admin panel unresponsive, blocking administrators from managing users.
- PyTorch Lightning denial of service vulnerability A malformed POST to the LightningApp state endpoint with unexpected values makes the server shut itself down.
- Quivr unauthenticated Denial of Service (DoS) via Multipart Boundary Appending characters after a multipart boundary makes the upload parser keep processing, tying up server resources without authentication.
- SageMaker Workflow component allows possibility of MD5 hash collisions Colliding MD5 hashes make different workflow configurations reuse each other's results, so workflows are silently replaced.
- Transformers Regular Expression Denial of Service (ReDoS) vulnerability Crafted input to the Nougat tokenizer's post-processing triggers catastrophic regex backtracking, pinning the CPU and stalling the process.
- vLLM allows Remote Code Execution by Pickle Deserialization via AsyncEngineRPCServer() RPC server entrypoints The RPC server passes incoming network messages directly to cloudpickle loads, so a crafted message executes code when it is deserialised.
- vLLM deserialization vulnerability in vllm.distributed.GroupCoordinator.recv_object A distributed communication helper unpickles received network bytes without checks, so a crafted message runs code on the receiving host.
- ZenML unauthenticated DoS via Multipart Boundry Extra characters appended to multipart boundaries send the parser into an infinite loop, hanging unauthenticated login requests.
- Flowise allows arbitrary file write to RCE A user-supplied file name joined to a storage path lacks validation, allowing writes outside the directory; overwriting package.json gives code execution.
- Flowise Pre-auth Arbitrary File Upload A path-prefix whitelist lets unauthenticated requests reach the attachments upload route, allowing arbitrary files to be written to local storage.
- PyTorch Tuple Handler is Vulnerable to Memory Corruption through Manipulation of None Argument Passing a None argument to a profiler callback helper causes out-of-bounds memory writes in the tuple handling code.
- PyTorch: Manipulation of the argument scale/zero_point leads to improper initialization via Quantized Sigmoid Module The quantized sigmoid in PyTorch leaves its output state improperly initialised when scale or zero_point arguments are manipulated.
- FlowiseAI Flowise arbitrary file upload vulnerability The attachments endpoint accepts arbitrary uploaded files without adequate type or content restrictions.
- Cybercriminals Reportedly Exploited Google’s G.Co Subdomain and Spoofed Caller ID in AI-Driven Phishing Attack on Hack Club Founder Attackers reportedly exploited Google's g.co subdomain and spoofed a Google Assistant number, with voice cloning and a genuine Workspace email, to phish Zach Latta.
- Label Studio allows Server-Side Request Forgery in the S3 Storage Endpoint A user-supplied S3 endpoint is passed unvalidated to the AWS SDK, so the server sends requests to internal hosts and echoes their responses in errors.
- Label Studio allows Cross-Site Scripting (XSS) via GET request to `/projects/upload-example` endpoint The GET upload-example endpoint renders HTML from a crafted label_config parameter without sanitisation, so opening the link runs attacker script.
- vLLM uses Python 3.12 built-in hash() which leads to predictable hash collisions in prefix cache Predictable Python 3.12 hash values let crafted prompts collide with cached prefixes, so a request can reuse cached state built from different content.
- Deep Java Library path traversal issue Archive extraction in DJL does not block absolute or cross-platform path entries, so a crafted model archive can write files outside the destination directory.
- Gradio Blocked Path ACL Bypass Vulnerability Blocked-path checks compare case-sensitively, so changing letter case gets past the access list on case-insensitive filesystems.
- keras Path Traversal vulnerability Extracting a crafted tar archive fetched by the file download helper writes files to paths outside the intended directory.
- khoj has an IDOR in subscription management allows unauthorized subscription modifications The subscription update endpoint looks up Stripe customers by a caller-supplied email and modifies their subscriptions without checking ownership.
- Letta (previously MemGPT) incorrect access control vulnerability The users endpoint of an agent memory server fails to enforce access control, exposing sensitive data to unauthorised callers.
- unstructured XML External Entity (XXE) Parsing a crafted XML document resolves external entities, letting the parser read local files or internal resources.
- @lobehub/chat Server Side Request Forgery vulnerability An unauthenticated chat endpoint accepts a user-supplied proxy address and sends requests to it, letting attackers reach internal services and echo responses back.
- MLflow's excessive directory permissions allow local privilege escalation Overly broad directory permissions let a local user win a race condition and escalate privileges while spark_udf runs.
- Flowise OverrideConfig security vulnerability Callers can override flow configuration at runtime, including prompts and node settings, reaching sandboxed code execution, SSRF, and data exfiltration.
- Missing ratelimit on passwrod resets in zenml The password change endpoint has no rate limit, so an attacker can brute-force the current password and take over the account.
- LightGBM Remote Code Execution Vulnerability A heap buffer overflow in LightGBM can be exploited to achieve remote code execution.