Browse / AI Framework and Platform Flaws / Other weaknesses
Other weaknesses
46 records in AI Framework and Platform Flaws.
- vLLM: Flash late-interaction scoring caches query embeddings under a caller-controlled request id — cross-request integrity break and induced errors on `/score` and `/rerank` A caller-chosen request ID lets a second request overwrite the cached query embedding, so documents are scored against the attacker's query.
- NLTK: Default ENFORCE=False Disables All pathsec Security Controls The security checks added for earlier CVEs only log warnings by default, so path traversal, unsafe URLs and unsafe pickle loading still proceed.
- New API: Integer overflow in quota billing yields negative charges (self-crediting) Oversized quantity fields wrap during integer conversion in billing, producing a negative quota that credits the attacker's balance at settlement.
- n8n: Cross-Tenant Module-Cache Poisoning in the JS Task Runner Code-node users poison a shared JavaScript module cache, altering other tenants' later Code-node executions on the same task runner.
- 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.
- GPTCache: File and image cache keys collide because BufferedReader.peek() only reads the buffered prefix Cache keys for file and image inputs use a weak hash over only the buffered prefix, so different files can collide and return each other's cached results.
- Gradio: Audio cache key ignores metadata when saving numpy audio outputs Gradio builds the audio cache key with a weak hash that ignores metadata, so colliding inputs can receive each other's cached audio.
- Open WebUI: Redis Cache Keys tool_servers and terminal_servers Missing Instance Prefix Enable Cross-Instance Cache Poisoning Redis keys for tool and terminal server configuration lack an instance prefix, so separate deployments sharing Redis overwrite each other's config.
- Claude Code: Insecure System-Wide Configuration Loading Enables Local Privilege Escalation on Windows A low-privileged Windows user creates the ClaudeCode folder under ProgramData and places a settings file that other users' sessions then load automatically.
- 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.
- 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.
- Flowise Allows Mass Assignment in `/api/v1/leads` Endpoint Object.assign copies every field from the unauthenticated request body onto the Lead entity, letting callers set internal fields such as id and chatId.
- Gradio has an Open Redirect in its OAuth Flow Logout and login callback endpoints redirect to an unvalidated target URL, letting attackers send users from a trusted domain to external sites.
- Vercel’s AI SDK's filetype whitelists can be bypassed when uploading files The AI SDK's file upload type allowlist can be circumvented, letting files of disallowed types be accepted.
- llama-index has Insecure Temporary File A shared, world-writable NLTK data directory lets local users replace or corrupt files that the library later loads.
- lobe-chat has an Open Redirect The OIDC redirect builds its target host from client-supplied forwarded headers, so an attacker can send users to an arbitrary domain.
- 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.
- 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.
- 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.
- docarray prototype pollution A dataset indexing function lets attacker-controlled keys modify object prototype attributes through a remotely reachable web API.
- BentoML Open Redirect vulnerability A crafted URL makes the BentoML server redirect visitors to an arbitrary external site.
- FastChat open redirect vulnerability A crafted URL parameter is used as a redirect target without validation, sending unauthenticated visitors to attacker-chosen external sites.
- Gradio Vulnerable to Open Redirect A URL-encoded crafted request makes the server return a 302 redirect to an attacker-chosen external site.
- 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.
- 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.
- 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.
- Gradio's dropdown component pre-process step does not limit the values to those in the dropdown list Crafted requests can send values outside the dropdown's configured list, ignoring the allow_custom_value restriction on the server side.
- Gradio has a race condition in update_root_in_config may redirect user traffic A race condition lets an attacker change the root URL the Gradio frontend uses, redirecting user requests and data to a malicious server.
- Open redirect in gradio Improper validation of user-supplied URLs lets attackers send victims through redirects to arbitrary external websites.
- Undefined Behavior in mlflow URL-encoded and decoded model names are treated as different models, letting an attacker register a same-named model that users later load by mistake.
- Race condition in zenml Parallel user creation requests with the same username both succeed, producing duplicate accounts and inconsistent user records.
- Clickjacking in zenml Missing X-Frame-Options and Content-Security-Policy headers let attackers embed the ZenML UI in a malicious iframe and trick users into unintended clicks.
- Gradio arbitrary file upload vulnerability Gradio's /upload route accepts arbitrary files without restricting type or destination.
- transformers has Insecure Temporary File Temporary files are created insecurely, so a local attacker sharing the system could access or tamper with them.
- Use of hard-coded, security-relevant constants in deepset-ai/haystack Haystack relies on hard-coded security-relevant constants, so its protections depend on fixed values that anyone reading the source can know.
- Invalid char to bool conversion when printing a tensor Tensor printing casts raw char values to bool, which is undefined behaviour for values other than zero or one and crashes sanitizer builds.
- Insecure Temporary File in mlflow Code creates temporary files with the deprecated mktemp function, which returns only a name and leaves a race window for attackers to plant files.
- Integer overflow in Tensorflow An unchecked multiplication of element count and element size in cost estimation can overflow into an incorrect size.
- Integer overflow in TensorFlow User-controlled crop parameters overflow an integer calculation in Grappler's cost estimator, triggering undefined behavior in TensorFlow.
- Integer overflow in Tensorflow An unchecked axis attribute in a quantization op's shape inference lets an attacker overflow an integer during rank calculation.
- Memory leak in decoding PNG images Invalid PNG input causes an early error return that skips freeing decoder buffers, leaking memory on each malformed image.
- Insecure temporary file in Tensorflow TensorFlow uses tempfile.mktemp, so another process can create the predicted filename first in a race and redirect the subsequent file operation.
- Reference binding to nullptr in shape inference Shape inference for SparseFillEmptyRows does not reject empty inputs, so it binds a reference to a null pointer, causing undefined behaviour.
- Reference binding to nullptr in unicode encoding Calling the UnicodeEncode op with empty split tensors makes the kernel bind a reference to null before checking for emptiness, causing undefined behaviour.
- Memory leak in Tensorflow Passing a list of strings to dlpack conversion leaks allocated memory because the error status from validation is never checked.