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
- BentoML Allows Remote Code Execution (RCE) via Insecure Deserialization Request payloads are unpickled without validation, so a crafted HTTP request executes arbitrary code on the server.
- PyTorch is Vulnerable to Memory Consumption through pad_packed_sequence Function A crafted input to the PyTorch packed sequence padding function causes memory corruption, but only with local access.
- PyTorch is vulnerable to memory corruption through its torch.lstm_cell function Crafted inputs to torch.lstm_cell cause memory corruption when the function is invoked locally.
- PyTorch is vulnerable to memory corruption through its torch.jit.script function A crafted input to torch.jit.script causes memory corruption in PyTorch, which can be triggered locally.
- PyTorch is vulnerable to memory corruption through its unpack_sequence function A crafted input to torch's unpack_sequence corrupts memory in native code, and the flaw can be triggered locally.
- GenNomis AI Database Reportedly Exposes Nearly 100,000 Deepfake and Nudify Images in Public Breach An unprotected database linked to the GenNomis image platform was publicly exposed, leaking nearly 100,000 nudify and deepfake files through a basic misconfiguration.
- PyTorch susceptible to local Denial of Service A local call to a max-pool operator with crafted input causes denial of service.
- LocalAI Cross-Site Scripting (XSS) vulnerability in its search functionality Search input is not properly sanitised before being rendered, letting injected JavaScript run in a victim's browser.
- Ollama Vulnerable to Denial of Service (DoS) via Crafted GZIP A malicious API server returns a gzip bomb that Ollama reads fully into memory, exhausting the server's memory.
- H2O Vulnerable to Arbitrary File Overwrite via File Export The H2O model export endpoint accepts any destination path, letting an attacker overwrite arbitrary server files with a model file.
- H2O Vulnerable to Denial of Service (DoS) via `HEAD` Request The typeahead endpoint makes an outbound HEAD request with no timeout, so a remote host that hangs can block the server from handling other requests.
- H2O Vulnerable to Denial of Service (DoS) via Large GZIP Parsing Repeatedly uploading and parsing a highly compressed gzip file amplifies data in memory, exhausting resources and stalling the server.
- H2O Vulnerable to Denial of Service (DoS) via `/3/ParseSetup` Endpoint A user-chosen regular expression is applied to attacker-controlled input in the parse setup endpoint, causing catastrophic backtracking that stalls the server.
- H2O Vulnerable to Arbitrary File Overwrite A user-controlled export directory parameter on the model details endpoint lets requests overwrite files at arbitrary paths on the host.
- H2O Deserialization of Untrusted Data Vulnerability Unauthenticated JDBC URLs passed to the import endpoints reach driver connection code, where untrusted deserialisation runs when a database driver is on the classpath.
- H2O Vulnerable to Execution of Arbitrary Files An exposed endpoint lets a caller encrypt arbitrary server files with a chosen key and overwrite that key, making recovery nearly impossible.
- H2O Vulnerable to Denial of Service (DoS) via `/3/ImportFiles` Endpoint The ImportFiles endpoint accepts a path that refers back to the same endpoint, so the server recursively calls itself until its request queue is exhausted.
- H2O Vulnerable to Denial of Service (DoS) via `/3/Parse` Endpoint A user-supplied string is compiled as a regular expression and applied to another string, so concurrent requests can exhaust server threads.
- AgentScope stored cross-site scripting (XSS) vulnerability A user-controlled run identifier is rendered into the run inspection page as raw HTML, so attacker script runs in viewers' browsers.
- AgentScope directory traversal vulnerability in /read-examples A crafted POST to the read-examples endpoint uses directory traversal to read any local JSON file on the server.
- AgentScope Cross-Origin Resource Sharing (CORS) vulnerability The server returns permissive cross-origin headers for any origin, so untrusted websites can read API responses from a visitor's browser.
- AgentScope path traversal vulnerability A delete-workflow endpoint uses a user-supplied path without restricting it to the intended directory, so files outside that directory can be removed.
- AgentScope Path Traversal in /api/file The file API endpoint uses an unsanitised path parameter, letting remote callers read arbitrary files on the server.
- AgentScope path traversal vulnerability in save-workflow Workflow save and load functions accept unvalidated paths, letting attackers read and write arbitrary JSON files on the host.
- BentoML Open Redirect vulnerability A crafted URL makes the BentoML server redirect visitors to an arbitrary external site.
- BentoML deserialization vulnerability Setting the args-number parameter above 1 on the BentoML runner server causes attacker-supplied data to be deserialised, executing arbitrary code.
- BentoML vulnerable to Uncontrolled Resource Consumption Appending characters to a multipart boundary in a request to the Gradio login route makes the server keep processing input until resources are exhausted.
- BentoML Denial of Service (DoS) via Multipart Boundary Appending dashes to a multipart boundary makes the parser keep processing characters, consuming resources until the service becomes unavailable.
- Feast Cross-Origin Resource Sharing vulnerability CORS settings accept any origin, so a third-party site can read API responses from a logged-in user's browser.
- FastChat open redirect vulnerability A crafted URL parameter is used as a redirect target without validation, sending unauthenticated visitors to attacker-chosen external sites.
- FastChat Denial of Service vulnerability A multipart upload with an oversized filename overwhelms the FastChat server, making it unavailable to other users.
- FastChat Server-Side Request Forgery vulnerability The web server lets requests reach internal addresses, so attackers can fetch internal resources such as cloud metadata credentials.
- FastChat Server-Side Request Forgery vulnerability Insufficient validation of a path parameter on the queue join endpoint lets requests be forwarded to internal hosts and cloud metadata services.
- FastChat Uncontrolled Resource Consumption vulnerability Extra characters after a multipart boundary cause the server's parser to loop indefinitely, exhausting resources for all users.
- Gradio DOS in multipart boundry while uploading the file A crafted multipart upload with a long run of extra characters after the boundary forces the parser to iterate per character and emit warnings, stalling the server.
- Gradio Vulnerable to Open Redirect A URL-encoded crafted request makes the server return a 302 redirect to an attacker-chosen external site.
- Gradio Vulnerable to Denial of Service (DoS) via Crafted Zip Bomb An uploaded compressed archive expands to a huge size when the dataframe component parses it with pandas, exhausting resources and crashing the server.
- Gradio Vulnerable to Arbitrary File Deletion The audio component builds file paths from user-controlled format values, letting an attacker truncate arbitrary files to empty.
- Gradio Path Traversal vulnerability The gradio blocked-path check fails to match NTFS alternate data stream syntax on Windows, so blocked files can still be read.
- Gradio Vulnerable to Denial of Service (DoS) via Crafted HTTP Request A crafted HTTP request to the Datetime component triggers catastrophic regex backtracking that pins server CPU at full load.
- H2O Vulnerable to Denial of Service (DoS) and File Write The run_tool command's ast-based parser exposes water.tools classes, one of which can stop the server and write large files to arbitrary directories.
- InvokeAI Arbitrary File Deletion vulnerability An image deletion API endpoint lets unauthorized callers delete arbitrary files on the server, including sensitive system and configuration files.
- InvokeAI has Denial of Service (DoS) vulnerability in `/api/v1/images/upload` Extra characters after a multipart boundary make the upload endpoint's parser loop forever, halting service for all users.
- InvokeAI Uncontrolled Resource Consumption vulnerability A very large board_name in a PATCH request makes the UI unresponsive and blocks board deletion.
- langchain-core allows unauthorized users to read arbitrary files from the host file system ImagePromptTemplate input variables can name arbitrary server file paths, which are read into prompts and can be exposed through the output.
- PyTorch Lightning path traversal vulnerability A crafted filename sent to the upload route on Windows lets a caller write files to arbitrary locations on the host.
- LiteLLM Vulnerable to Remote Code Execution (RCE) A config value for a post-call callback is split into module and function names and imported, letting an attacker point it at os.system.
- LiteLLM Vulnerable to Denial of Service (DoS) via Crafted HTTP Request Trailing dash characters appended to a multipart boundary make the proxy's parser loop over each character, exhausting resources without authentication.
- LiteLLM Has an Improper Authorization Vulnerability A viewer role receives an overly privileged API key, letting it call admin endpoints and promote itself to proxy admin.
- LiteLLM Vulnerable to Denial of Service (DoS) Unauthenticated requests reach an ast.literal_eval call on user-supplied input, which can exhaust resources and crash the LiteLLM server.