AI Security
New Attack Method Exploits LLM Agents to Drain SaaS Provider Resources
Researchers reveal how malicious users can craft deceptive prompts and tool configurations to cause AI agents to repeatedly waste computing resources on SaaS platforms.
A newly disclosed vulnerability in software-as-a-service platforms powered by large language model agents could allow attackers to dramatically amplify resource consumption on provider infrastructure with minimal effort.
The Core Vulnerability#
The attack exploits a fundamental assumption in subscription-based SaaS deployments: that the AI agent itself decides when to stop executing tasks. In such systems, both model and tool calls consume provider resources without triggering additional charges to the customer. A malicious user can send a specially crafted prompt that tricks the agent into repeatedly performing work long after the original task should have finished—burning through computing resources without paying for the extra usage.
How OverServe Works#
Researchers have published a framework called OverServe that demonstrates this attack pattern. The technique works by co-designing two components: a fake status tool (built using the Model Context Protocol) and a prompt designed to exploit the agent's decision-making.
The malicious tool reports accurate status information and returns without errors, yet the accompanying prompt causes the agent to repeatedly check whether a condition has been met. Since the tool never signals completion, the loop continues indefinitely—forcing the agent to make unnecessary API calls and consume resources even after the legitimate business task is done.
Testing and Results#
The research team tested the attack against several prominent LLM models, including Qwen, Nemotron, DeepSeek, and Claude Sonnet, using tasks from nine different benchmarks. The findings showed:
- All OverServe-generated tool specifications passed existing MCP security scanners without raising alerts
- Every attack run successfully completed the original task while also consuming extra resources
- The amplification effect reached a maximum of 93.88× more API requests than necessary
Why This Matters#
This vulnerability highlights a structural risk in how AI agents are deployed in revenue-critical environments. Providers cannot easily distinguish between legitimate work and resource-wasting loops without fundamentally changing how agents operate. The attack is effective precisely because it doesn't break any explicit rules—the agent is simply following its instructions to completion.

