What happened
Published ~2026-11-13, the OverServe paper (arXiv:2611.14808) presents a multi-agent framework that combines an MCP status-tool and a completion-confirmation prompt so that a malicious SaaS tenant can force an agent to keep 'working' on an already-finished task. Because the tool returns accurate, normally-terminating status and each component looks benign, the approach evades the detection/auditing that defeats single-vector prompt-only or tool-only resource-amplification attacks.
Why it matters
For AI deployments, this is a novel agent-execution attack class that converts the subscription economics against the provider: a single malicious tenant can quietly multiply the provider's LLM and tool-call bill while preserving task correctness and avoiding MCP scanner alerts — the agentic equivalent of a denial-of-wallet / cost-amplification attack, with a working framework and measured 93.88x amplification.
Attack vector
A malicious tenant acts as a normal SaaS user and supplies a benign-looking prompt plus an attacker-controlled MCP 'status' tool. The tool returns truthful, always-'pending' status; the agent keeps re-checking an unfinished condition even after the real business task completed, consuming the provider's tokens/tool budget until the execution budget is exhausted — a denial-of-wallet / resource-amplification attack. Up to 93.88x amplification of API requests while still 'passing' the original task checks and triggering no alerts in the evaluated MCP scanners.
Affected systems
SaaS LLM-agent deployments using Model Context Protocol (MCP) tools; generalizes across agentic SaaS in CRM/ERP/security verticals (no single vendor version range — an attack-methodology class).
Mitigation
For agentic SaaS operators: bound agent execution steps/token budgets per request, require human-in-the-loop confirmation on long-running or stateful tasks, treat third-party MCP status tools as untrusted, and add resource/cost guards that stop an agent from polling a never-completing condition. (No CVE; academic framework from arXiv:2611.14808.)