Researchers Demonstrated AI Agent-Tool Fan-Out Risks

Forcepoint simulated how poisoned inputs can cause excessive resource consumption in autonomous AI agents.

Updated on Sept. 22, 2026 in Artificial Intelligence

Isometric editorial illustration of a complex cluster of server pipes and cabling, representing systemic resource exhaustion vulnerabilities in AI.
Security firm Forcepoint revealed that autonomous AI agents are vulnerable to recursive tool call attacks that can lead to resource exhaustion and excessive operating costs. AI Illustration. Upload story photo >

Live Poll

Do you trust companies to effectively manage the security risks associated with their AI tools?

Security firm Forcepoint has demonstrated a potential vulnerability where poisoned AI agent inputs trigger recursive tool calls. This agent-tool fan-out attack can lead to significant cost spikes and resource exhaustion.

Why it matters

Autonomous agents process accumulating context that increases per-turn costs, making them vulnerable to exploits that force excessive operations. These attacks risk "denial-of-wallet" scenarios where malicious actors intentionally inflate operating expenses for organizations.

Agents with recursion limits and circuit breakers can complete calls for $0.02, while unrestricted agents may exceed $10 per poisoned source. Simulation results show per-turn costs scaling from $0.001 to $0.50 by turn 100.

The players

Forcepoint

Forcepoint is a cybersecurity firm that conducts research into emerging threats and vulnerabilities within enterprise-grade AI agent systems.

OWASP

The Open Web Application Security Project is a non-profit foundation that tracks and ranks the most critical security risks facing web and AI applications.

The details

Attackers can poison an AI agent's input, causing it to engage in excessive reasoning loops or repetitive tool calls that deplete computational resources. These vulnerabilities include reasoning-loop exhaustion and model extraction attacks, where attackers use high query volumes to replicate targeted models.

Timeline

  1. In 2025, unbounded consumption ranked No. 10 on the OWASP list.

  2. By 2026, the unbounded consumption risk rose to No. 6 on the OWASP list.

The Tech Race

The rising threat of unbounded consumption signals a shift toward prioritizing financial and operational safety as AI agents move from experimental prototypes to enterprise deployments. This vulnerability mirrors the early security challenges faced by cloud computing architectures during their initial scaling phase.

Developers and organizations must implement recursion depth limits and circuit breakers to prevent unforeseen operational cost spikes. These security measures ensure that rogue or poisoned agents cannot trigger infinite loops that drain budget and processing power.

The takeaway

Organizations should treat agent-based AI integrations as high-risk environments requiring strict cost-capping protocols. Proactively setting limits on tool calls and turn counts can shield developers from malicious attempts to trigger resource depletion.

Further reading

Learn more about the security implications of autonomous systems in our Artificial Intelligence section.

Source note: This article includes information reported by SC Media.

Live Poll

Do you trust companies to effectively manage the security risks associated with their AI tools?