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
- Hugging Face Transformers library has Regular Expression Denial of Service A regular expression in the English number normaliser backtracks catastrophically on long digit strings, consuming CPU until the service stalls.
- Hugging Face Transformers is vulnerable to ReDoS through its MarianTokenizer A regular expression in the MarianTokenizer language code removal takes excessive time on crafted strings, pinning the CPU.
- Flowise Cloud and Local Deployments have Unauthenticated Password Reset Token Disclosure that Leads to Account Takeover The unauthenticated forgot-password endpoint returns a valid reset token in its response, letting an attacker reset any account's password by email address.
- N8N's Chat Trigger component is vulnerable to XSS The Chat Trigger accepts arbitrary file uploads, including HTML files that execute script when opened in the application's origin.
- Langchain Community Vulnerable to XML External Entity (XXE) Attacks The Evernote loader parses attacker-supplied XML with external entities enabled, letting it read local files such as system password files.
- LlamaIndex affected by a Denial of Service (DOS) in JSONReader Deeply nested JSON input causes unbounded recursion in the reader, exceeding Python's recursion limit and exhausting resources or crashing the process.
- Langflow Vulnerable to Privilege Escalation via CLI Superuser Creation (Post-RCE) An attacker who already has code execution in the container runs the superuser CLI command, which creates an admin account without checking for an existing one.
- vllm API endpoints vulnerable to Denial of Service Attacks An oversized HTTP header on a GET request makes the server buffer the entire value in memory, exhausting available RAM.
- n8n symlink traversal vulnerability in "Read/Write File" node allows access to restricted files A user who can create symlinks, for example through a command node, can point the file node past its directory restrictions to read or write protected paths.
- Stored XSS in n8n Form Trigger allows Account Takeover via injected iframe and video/source An authenticated user stores HTML in a form trigger using iframe srcdoc or video and source onerror payloads, which run in victims' browsers and steal session tokens.
- Ollama allows deletion of arbitrary files A crafted request to the model pull endpoint makes the server delete files at attacker-chosen locations outside the intended model store.
- Hugging Face Transformers Regular Expression Denial of Service (ReDoS) vulnerability A catastrophic-backtracking regex in the TensorFlow-to-PyTorch weight name converter lets crafted names consume excessive CPU during conversion.
- BentoML SSRF Vulnerability in File Upload Processing The framework automatically downloads URLs supplied in multipart or JSON file fields without checking internal addresses, so callers can make the server request internal services.
- Reported Hack of Tea Dating App Compromises Data from Purportedly AI-Supported Identity and Image Checks Attackers breached a legacy storage system of the Tea app, exposing about 72,000 users' selfies and photo IDs; the cause was an ordinary storage breach, not an AI flaw.
- Ollama vulnerable to Cross-Domain Token Exposure A malicious WWW-Authenticate realm makes the client send its stored bearer token to an attacker-chosen host during a registry pull.
- Transformers is vulnerable to ReDoS attack through its DonutProcessor class Crafted input strings trigger catastrophic regex backtracking in the Donut token-to-JSON parser, pinning CPU during document processing.
- LlamaIndex vulnerable to Path Traversal attack through its encode_image function Traversal sequences in the image path argument let callers read arbitrary local files through the image encoding helper.
- LlamaIndex vulnerable to DoS attack through uncontrolled recursive JSON parsing Deeply nested JSON passed to the reader causes unbounded recursion that exhausts the stack and crashes the process.
- Lord of Large Language Models vulnerable to Observable Discrepancy attack via authenticate_user function Password comparison exits at the first mismatched character, so response timing reveals valid usernames and lets attackers guess passwords incrementally.
- Transformers vulnerable to ReDoS attack through its SETTING_RE variable A regex used by the chat command has nested quantifiers, so crafted near-matching input causes exponential backtracking and stalls the process.
- Transformers vulnerable to ReDoS attack through its get_imports() function A catastrophic backtracking regex in the import-scanning helper makes crafted Python source consume excessive CPU when model code is loaded.
- Transformers's Improper Input Validation vulnerability can be exploited through username injection A prefix-based URL check can be fooled by placing a fake YouTube host in the username part of the URL, so malicious domains pass validation.
- Transformers's ReDoS vulnerability in get_configuration_file can lead to catastrophic backtracking A crafted configuration filename triggers catastrophic regex backtracking in the config lookup, consuming CPU and stalling model loading or serving.
- Microsoft 365 Copilot Vulnerability Allegedly Allowed File Access Without Audit Log Entry A Microsoft 365 Copilot flaw let users access and summarize files without generating audit log entries, and Microsoft fixed it without notifying customers or assigning a CVE.
- n8n is vulnerable to Improper Authorization through its `/stop` endpoint The execution stop endpoint does not check workflow ownership, so an authenticated user can halt runs belonging to other users.
- n8n Vulnerable to Denial of Service via Malformed Binary Data Requests Requests to the binary-data endpoint with empty filesystem URIs trigger resource exhaustion, letting an authenticated user make the service unresponsive.
- n8n allows open redirects via the /signin endpoint The post-login redirect parameter on the sign-in page is not restricted to same-origin paths, so users can be sent to attacker domains.
- MLFlow SSRF via gateway_proxy_handler The gateway proxy handler forwards requests with an unvalidated path, letting callers make the server reach arbitrary internal URLs.
- Langflow Unauth RCE An unauthenticated code validation endpoint passes attacker-supplied code into execution, giving remote code execution over HTTP.
- Gradio Allows Unauthorized File Copy via Path Manipulation The flagging feature copies any server file named in an unauthenticated predict request, without validating the supplied path first.
- Gradio CORS Origin Validation Bypass Vulnerability A flaw in localhost alias handling in the CORS origin check lets a disallowed origin pass validation and read responses from the server.
- LLama-Index CLI OS command injection vulnerability The CLI passes its files argument unescaped into os.system, so shell metacharacters in a filename run arbitrary commands.
- vLLM DOS: Remotely kill vllm over http with invalid JSON schema An invalid JSON schema in a guided decoding request causes an unhandled exception in the grammar compiler that crashes the engine.
- vLLM allows clients to crash the openai server with invalid regex A client-supplied invalid regex in structured output makes the vLLM OpenAI server crash.
- vLLM has a Weakness in MultiModalHasher Image Hashing Implementation The image hash uses only raw pixel bytes without dimensions or mode, so different images can share a hash and receive each other's cached results.
- vLLM vulnerable to Regular Expression Denial of Service Several regular expressions in vLLM backtrack catastrophically on crafted input, so a single request can stall processing.
- docarray prototype pollution A dataset indexing function lets attacker-controlled keys modify object prototype attributes through a remotely reachable web API.
- vLLM Allows Remote Code Execution via PyNcclPipe Communication Service The PyNcclPipe KV transfer service accepts pickled messages from a TCP store bound to all interfaces, so remote senders can execute code on the server.
- Hugging Face Transformers Regular Expression Denial of Service Nested quantifiers in a docstring code-block regex cause catastrophic backtracking on inputs with many newlines, consuming CPU.
- label-studio vulnerable to Cross-Site Scripting (Reflected) via the label_config parameter. A crafted label_config posted to the upload-example endpoint is reflected into a response the browser renders, running script in the victim's session.
- LlamaIndex Vulnerable to Denial of Service (DoS) Recursive link crawling in the web reader lacks an effective max_depth limit, so repeated calls exhaust Python's recursion limit and crash the process.
- Serviceaide AI Platform Implicated in Health Data Exposure Affecting 483,000 Catholic Health Patients A misconfigured Elasticsearch database in Serviceaide's agentic AI platform exposed medical records, insurance details and credentials of 483,000 Catholic Health patients.
- Remote Code Execution Vulnerability in vLLM Multi-Node Cluster Configuration Secondary vLLM hosts unpickle messages received over ZeroMQ, so anyone who controls the primary host or its network can run code on them.
- Transformers Regular Expression Denial of Service (ReDoS) vulnerability A backtracking regex in the Japanese GPT-NeoX tokenizer takes exponential time on crafted text, pinning CPU and stalling the service.
- Data exposure via ZeroMQ on multi-node vLLM deployment A multi-node vLLM host binds an internal publish socket to all interfaces, so any reachable client receives broadcast internal state.
- vLLM Vulnerable to Remote Code Execution via Mooncake Integration The Mooncake KV transfer integration calls recv_pyobj over a ZeroMQ socket listening on all interfaces, so attacker-sent pickle data runs code when deserialised.
- n8n Vulnerable to Stored XSS through Attachments View Endpoint The attachment view endpoint lets callers choose the response MIME type, so an uploaded HTML file can be served as text/html and run script.
- one-api Cross-site Scripting vulnerability Homepage content set through system settings is rendered without sanitisation, so injected script runs in visitors' browsers.
- PyTorch Improper Resource Shutdown or Release vulnerability A local caller can trigger improper resource release in the CTC loss function, causing the process to stall or crash.
- FlowiseDB vulnerable to SQL Injection by authenticated users An authenticated user imports a crafted chatflow file whose ID contains traversal and SQL-breaking input, which the import routine concatenates into a raw SQL query.