Latest
●Prompt injection turns up in website HTTP headers, study finds●Nezha dashboard can be crashed by any low-privilege member●MCP Python SDK kept sessions forever, letting memory run out●Nginx UI flaws let MCP tools skip config security checks●Bank of America adds 1,000 contractors to fix Mythos-found bugs●UK stops short of AI rules in Cyber Security Bill●GhostAction campaign hits 1,900 GitHub accounts and 30,000 repos●Attackers target critical Langflow flaw in 1,423 exploit attempts
Friday, 16 October 2026
1 agent hack today▼ 7 vs yesterday (8)
The Agentic TimesPowered byVortal
Explainers

Runaway agent spending flagged as an enterprise risk

Dark Reading reports that unbounded consumption by AI agents can drive unexpectedly large bills, an issue OWASP ranks sixth in its LLM Top 10.

By The Agentic Times ·

A mechanical spider spins an endless web of golden thread across a factory floor, the pile growing faster than workers can contain it, threatening to collapse the warehouse.
A mechanical spider spins an endless web of golden thread across a factory floor, the pile growing faster than workers can contain it, threatening to collapse the warehouse. · Illustration: The Agentic Times

Enterprises running AI agents face a cost risk as well as a security one, according to a Dark Reading article published on 21 September 2026 examining how AI agents can trigger runaway costs.

The piece focuses on unbounded consumption, which the Open Worldwide Application Security Project (OWASP) currently ranks sixth in its Top 10 for LLM Applications, a list of the most significant risks facing systems built on large language models. According to Dark Reading, the problem could prove extremely costly for organisations that deploy agents without limits on what those agents are allowed to spend or consume.

The ranking matters because the OWASP Top 10 for LLM Applications has become a common reference point for security teams deciding what to test for and what controls to put in place. Its inclusion of unbounded consumption places a resource and billing problem alongside more familiar security concerns in the same list. For organisations that treat the OWASP list as a checklist, that means budget controls sit within the security remit rather than purely with finance or procurement.

The risk is particular to agentic systems because agents make repeated calls to models and tools on their own, rather than waiting for a person to click a button each time. Each call carries a cost. Where a traditional application's usage broadly tracks the number of users, an agent's usage tracks the number of steps it decides to take, and that figure is far harder to predict in advance.

Dark Reading's article is framed as guidance for enterprises rather than a report of a specific breach or a disclosed vulnerability. It does not describe a single named incident. The practical question it raises for security and platform teams is whether their agent deployments have hard ceilings such as rate limits, spending caps and alerting, and whether anyone is watching consumption in something close to real time.

For teams reviewing their own exposure, the OWASP ranking offers a starting point: if unbounded consumption is judged one of the ten most significant risks to LLM applications, an agent that can call models or paid third-party services without a defined stopping condition is running without one of the controls the industry's most widely cited list expects to see.

Sources

  1. darkreading.comhttps://darkreading.com/application-security/how-ai-agents-can-trigger-runaway-costs