Browse / Unbounded Consumption
Unbounded Consumption
Denial of service, resource exhaustion and cost attacks. 50 records. Reference: OWASP LLM10 · ATLAS AML.T0029.
- vLLM: DoS via unbounded chat_template Jinja2 rendering exhausts the renderer thread pool A single short chat template with unbounded Jinja2 rendering runs in a limited worker pool that cannot be interrupted, blocking chat endpoints for all tenants.
- vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine A legal maximum-length audio clip makes the Phi-4 audio encoder keep full attention matrices for every layer in memory, exhausting GPU memory and crashing the engine.
- vLLM DeepSeek-OCR image processing vulnerable to denial of service A valid image whose normalised pixels sum to exactly zero hits an inherited early-return guard in the DeepSeek-OCR path, crashing the serving process.
- vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash / DoS) A pre-tokenized prompt with an id one past the embedding range passes input validation and crashes the inference engine's CUDA context.
- vLLM: Responses API bypasses trust_request_chat_template for chat_template_kwargs The Responses endpoint accepts request-supplied Jinja chat templates that skip the trust-flag check, exhausting CPU and memory.
- vllm p2p never connected zmq_exhaustion Repeated API requests naming many unreachable peer addresses leave ZeroMQ sockets open, exhausting the socket quota and crashing the inference engine.
- vLLM : DoS via unvalidated mm_processor_kwargs in multimodal image processors (Isaac, InternVL, H2OVL, etc.) Clients can set per-request processor arguments such as image size that are passed unchecked to processors, forcing huge allocations that exhaust server memory.
- vLLM: Forged multimodal UUID + P0/P1 processor-cache eviction drift: unauthenticated remote crash of the whole vLLM engine Attacker-chosen multimodal UUIDs desynchronize two processor caches, so a crafted image request crashes the whole inference engine without authentication.
- Pydantic AI: Concurrency-limited models can keep their slot when a streamed request ends early Streams that end early or are consumed from another task fail to release shared concurrency slots, so repeated disconnects gradually block other requests.
- Pydantic AI: Event loop blocked by quadratic title extraction in `web_fetch` An attacker who can steer the agent's web fetch tool to a crafted page causes quadratic processing that blocks the event loop.
- Pydantic AI: Unbounded memory use when downloading remote content via web_fetch or FileUrl An agent's web-fetch tool buffers the entire response from a model-chosen URL before size limits apply, so a large body exhausts worker memory.
- Pydantic AI: Excessive resource use when local web fetching converts nested HTML Attacker-controlled HTML fetched by a Pydantic AI agent expands during conversion before size limits apply, consuming large amounts of CPU and memory.
- vLLM: Loose `cache_salt` validation lets a single request kill EngineCore on LMCache-MP deployments — uncaught downstream `ValueError` denial of service One request with a cache_salt containing a forbidden character passes loose validation and raises an uncaught error that kills the vLLM engine process.
- vLLM: GLMGA video sampling permits request-driven CPU and memory exhaustion Chat requests can set video sampling options that make the server build a huge frame index, consuming CPU and memory and delaying other media requests.
- vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach Unauthenticated requests set uncapped frame and fps limits for Qwen vision models, inflating API server memory use.
- vLLM: Request-selected PyNvVideoCodec GPU decode bypasses static VRAM reservation A client can pick a GPU video decoder per request, allocating GPU memory the engine never reserved and exhausting VRAM.
- vLLM: Unauthenticated audio decompression-bomb DoS in /v1/chat/completions An unauthenticated client sends a small compressed audio clip to the chat endpoint, which expands to gigabytes on decode because the duration limit applies only to transcription.
- vLLM: ReDoS via structured_outputs.regex in the lm-format-enforcer backend (no compile timeout) — missed sibling of GHSA-rwxx-mrjm-wc2m A structured output request with a catastrophic regex is compiled without a timeout in the lm-format-enforcer backend, stalling the inference worker.
- vLLM: Derender endpoints decode caller-supplied GenerateResponse token IDs without output bounds Derender routes detokenize client-supplied token ID lists without output size or count limits, consuming CPU and memory.
- New API: Redis user quota cache overwrite via PUT /api/user/self allows quota bypass A settings update writes a stale full user snapshot back to Redis, overwriting concurrent quota deductions so users can keep making model calls beyond their paid quota.
- vLLM: Completion prompt lists fan out into unbounded engine requests A completions request with a list of prompts is expanded into one engine generator per element with no outer count limit, so one request spawns many backend subrequests.
- Open WebUI: Any authenticated user can stall a worker via a knowledge-search pattern that backtracks catastrophically A chat user can get the model to run a catastrophic-backtracking regex against knowledge files, and no timeout stops it from pinning a worker.
- vLLM denial of service via prompt embeds on M-RoPE models A prompt-embeds request to completions on an M-RoPE model trips an assertion that crashes the engine and the whole server.
- vLLM has Remote DoS via Invalid Recovered Token Reinjection An overlapping generate and abort sequence makes speculative decoding feed an out-of-vocabulary sentinel token back into the drafter, crashing the shared worker.
- vLLM: ReDoS via structured_outputs.regex compiled without timeout in xgrammar and outlines backends A user-supplied regex for structured output is compiled by grammar backends with no timeout, so a catastrophic pattern hangs the inference worker.
- vLLM: OOM Denial of Service via Audio Decompression Bomb A small compressed audio upload decodes into a very large float buffer, exhausting server memory because the size check only applies to compressed bytes.
- vLLM: temperature=NaN and temperature=Infinity bypass validation and propagate to GPU kernels Temperature checks use comparisons that let NaN and positive infinity through, so these values reach GPU sampling kernels and crash the inference worker.
- vLLM is vulnerable to an Out-of-Memory (OOM) Denial of Service (DoS) attack due to unbounded frame count processing in the `VideoMediaIO.load_base64()` method A single unauthenticated chat completion request with a data URL containing thousands of comma-separated JPEG frames forces the server to decode them all into memory.
- Open WebUI: Unauthenticated endpoint can trigger embedding generation (cost/DoS) An endpoint that runs the embedding function needs no login, so anonymous callers can trigger repeated embedding computation and paid provider calls.
- OpenAI-Linked AI Agents May Have Contributed to a Partial Outage of the Wikimedia Foundation's Data Query Service Wikimedia reported AI agents linked to OpenAI generated heavy traffic and unauthorized edits, contributing to a partial outage of its Wikidata query service.
- vLLM Vulnerable to Remote DoS via Special-Token Placeholders Unauthenticated prompts spelling special placeholder tokens without matching media cause an out-of-range index that crashes the worker.
- FastChat has Denial of Service Through Blocking Event Loop in Model Workers (Incomplete Fix for ff66426) Long-running generation requests in model worker APIs block the event loop, and an earlier fix missed several endpoints, so other routes still stall service.
- PraisonAI: Unauthenticated WebSocket Endpoint Proxies to Paid OpenAI Realtime API Without Rate Limits An unauthenticated client can open unlimited WebSocket sessions that each spend the server's OpenAI key with no message or size limits.
- vLLM: Unauthenticated OOM Denial of Service via Unbounded `n` Parameter in OpenAI API Server An unauthenticated request with an enormous n value makes the API server copy request objects until it blocks the event loop and runs out of memory.
- vLLM: Denial of Service via Unbounded Frame Count in video/jpeg Base64 Processing The base64 video/jpeg path splits frames without applying the frame limit, so one request with thousands of frames exhausts server memory.
- Hugging Face Text Generation Inference vulnerable to Uncontrolled Resource Consumption Remote image links in prompts are fetched in full and buffered in memory during validation, even when the request is later rejected for size.
- vLLM is vulnerable to DoS in Idefics3 vision models via image payload with ambiguous dimensions A 1x1 image is misread as channel-first, causing a tensor split mismatch in the Idefics3 vision path that crashes the engine process.
- vLLM introduced enhanced protection for CVE-2025-62164 Crafted sparse prompt embeddings reach PyTorch without index validation, so enabling the feature lets a user crash or stall the serving process.
- vLLM vulnerable to DoS via large Chat Completion or Tokenization requests with specially crafted `chat_template_kwargs` Caller-supplied chat template keyword arguments can flip the tokenize default, forcing blocking tokenisation of huge inputs that stalls the API server.
- vLLM vulnerable to DoS with incorrect shape of multimodal embedding inputs Multimodal embedding inputs with the right number of dimensions but a wrong hidden size pass validation and crash the engine during embedding scatter.
- vLLM: Resource-Exhaustion (DoS) through Malicious Jinja Template in OpenAI-Compatible Server Malicious Jinja templates, or keyword arguments that overwrite the template key, exhaust CPU and memory during rendering, bypassing the chat_template restriction.
- AI-Powered Taco Bell Drive-Thru Reportedly Disrupted by Viral Prank Ordering 18,000 Water Cups A prankster deliberately flooded the voice-based drive-thru AI by ordering 18,000 water cups, overwhelming the system and forcing staff to intervene.
- vLLM Tool Schema allows DoS via Malformed pattern and type Fields Malformed type and pattern fields in a tool schema reach the grammar and regex converters unvalidated, crashing the inference worker until it is restarted.
- vLLM has a Regular Expression Denial of Service (ReDoS, Exponential Complexity) Vulnerability in `pythonic_tool_parser.py` A nested regex used to detect tool calls in generated text backtracks catastrophically on crafted output, stalling the serving process.
- vLLM: Quadratic Time Complexity in Input Token Processing leads to denial of service Repeated placeholder token expansion rebuilds the input list each time, so crafted inputs with many placeholders consume quadratic CPU time.
- vLLM vulnerable to Denial of Service by abusing xgrammar cache Repeated short structured-output requests with unique schemas fill the grammar cache until the server runs out of memory.
- vLLM denial of service via outlines unbounded cache on disk Many short requests with unique structured-output schemas make vLLM's outlines backend grow an on-disk grammar cache until the disk fills.
- LLM Scrapers Allegedly Target Multiple Open Source Projects Disrupting the FOSS Ecosystem Aggressive AI scrapers from Alibaba IP ranges ignored robots.txt and spoofed browser headers, overwhelming KDE, GNOME and other FOSS sites and causing outages.
- vLLM Denial of Service via the best_of parameter Very large best_of values in vLLM completion requests consume excessive resources without timeouts, leaving the API unresponsive to other users.
- Alleged LLMjacking Targets AI Cloud Services with Stolen Credentials Attackers used stolen cloud credentials, obtained through a vulnerable Laravel server (CVE-2021-3129), to run costly requests against Claude and AWS Bedrock and resell access via reverse proxies.