Unbounded Consumption Misconfiguration

Without proper limits on the number of queries or processing time, LLM APIs can become overwhelmed, affecting the stability and availability of the system, potentially resulting in denial-of-service attacks or resource depletion.

Unbounded consumption occurs when LLM APIs are not adequately rate-limited, throttled, or otherwise restricted in terms of resource use. This can involve excessive model inference requests, unbounded query sizes, or long-running processes that put undue strain on the server hosting the model, causing delays, outages, or even a complete loss of service.

When LLM APIs are not properly controlled, an attacker can abuse the system by sending high volumes of requests, large data queries, or complex tasks that drain resources. This issue can compound in multi-tenant systems where excessive consumption from one user or process affects others, leading to potential performance bottlenecks or system downtime.
Additionally, a Denial of Wallet (DoW) attack can occur when attackers exploit the cost-per-use model of cloud-based AI services by initiating a high volume of requests, which leads to unsustainable financial burdens on the provider. This can result in financial ruin for the provider, as the cost of processing excessive requests becomes unmanageable.

Example 1: A user submits a series of large text generation requests to an LLM API without any timeout, which allows requests to run without limits. This causes the system to process excessive data and exhausts available server resources.

import openai

# OpenAI GPT model setup
openai.api_key = "your-api-key"

def generate_large_response(prompt):
  response = openai.Completion.create(
      engine="text-davinci-003",
      prompt=prompt,
      max_tokens=1000
  )
  return response

# Submitting unbounded requests with high complexity and token count
prompt = "Generate an extensive article on AI."
response = generate_large_response(prompt)
print(response)

Example 2: A chat model is utilized without configuring a rate_limiter parameter while interacting with the LLM model.

from langchain_anthropic import ChatAnthropic
model = ChatAnthropic(
	model_name="claude-3-opus-20240229",
# not using rate_limiter
)

These examples demonstrate how unbounded API calls that lack a timeout can consume significant resources. Without restrictions on the execution time, the system can become overloaded. This can negatively affect the stability of the service and possibly result in resource exhaustion or denial-of-service risks.