ExSitu

An Accounting Agent "Ran Amok" and Racked Up $50,000 in an Hour — What the Mandiant Report Reveals

The AI Risk and Resilience report from Mandiant and Google GTIG reveals an AI agent that made 15,000 API calls in an hour, as well as the first zero-day attack created by AI.

5 min read

An Accounting Agent "Ran Amok" and Racked Up $50,000 in an Hour — What the Mandiant Report Reveals
Unsplash · Tyler

On September 16, 2026, Mandiant and the Google Threat Intelligence Group (GTIG) published the 2026 AI Risk and Resilience report — a look at how AI agents are already changing both defense and attack in organizations.

The $50,000 Incident

An AI agent responsible for accounting ran amok and made more than 15,000 expensive API calls within a single hour, at a cost of about $50,000, while disrupting live business transactions. There was no external attacker here — just an agent that entered an uncontrolled execution loop with no budget limit.

The Leading Attack Vectors

Prompt injection remains the dominant vector in enterprise AI applications. Alongside it, the report documents supply chain attacks: in February 2026 VirusTotal identified malicious OpenClaw packages posing as legitimate automation, and in March the threat group UNC6780 exploited supply chain weaknesses to steal access credentials to AI systems using prompt injection.

The Most Worrying Point

In May 2026, GTIG disclosed for the first time a verified case of an attacker using an AI-developed zero-day exploit in a campaign planned for mass distribution. It is the first time a major threat intelligence body has confirmed such an incident as fact, not as a future scenario.

What the Report Recommends

Three main recommendations: moving to clear, adaptive identity controls instead of blanket permissions; accelerating defensive response so it approaches the speed of attack; and refocusing the security operations center (SOC) on real-time behavioral telemetry instead of relying exclusively on known signatures.

What This Means for Your Organization

  • The central story in the report is an accounting agent that ran amok and made more than 15,000 expensive API calls in just one hour, at a cost of about $50,000, while disrupting live business transactions at that very moment. This is not a theoretical scenario from the professional literature — it is what actually happens when an AI agent is given access to systems without a clear budget limit set in advance.
  • The immediate practical lesson: for every AI agent that performs actions with a real cost — paid API calls, transactions, automated purchasing — set a hard daily spending cap and a person who approves any overrun, exactly as you would with a new employee holding an open corporate credit card.
  • According to the report, prompt injection is still the dominant vector in enterprise AI applications. If you have not yet tested your agents and chatbots against targeted injection scenarios, now is the time — the attackers have already done it for you.
  • The report also documents supply chain attacks: AI automation tools and add-ons that posed as legitimate software and contained a hidden backdoor. Before installing any external add-on or "skill" for an AI agent, check its source and developer exactly as you would check a new code library before deploying it to production.
  • In May 2026, the Google Threat Intelligence Group documented for the first time a verified case of an attacker using an AI-developed zero-day exploit in a planned attack campaign. It is a sign that the gap between "theoretical" and "happening in the field" is closing faster than most organizations are currently prepared for.
  • The report's main recommendation — moving to real-time behavioral monitoring instead of relying exclusively on perimeter defense — fits a small or mid-sized organization too: know not only what the agent is supposed to do according to its configuration, but what it is actually doing at any given moment, and who is checking.

Sources