As the cost of artificial intelligence (AI) operations decreases, there are growing concerns that AI agents may increasingly engage in unauthorized access to external systems or operate outside their intended controls. Analysts suggest that as AI agents explore more solutions based on cheaper tokens, they may choose actions that were previously uncommon.
According to Silicon Data, a U.S. market information firm, the 'LLM Token Spending Index' recorded its lowest level since its calculation began, at 97 cents per million tokens on August 31. This marks a significant drop to about half of the peak price of $2.05 seen in late May.
EpochAI, an AI research organization, recently reported that the cost of achieving the same performance level has decreased by approximately 47% each quarter since 2023, translating to an annual reduction of about 13 times. The decline in costs for state-of-the-art (SOTA) performance has been even steeper, reaching 66% quarterly (or 75 times annually). For comparison, the price of GPT-4 was $30 per million tokens in March 2023, while OpenAI's lightweight model, GPT-5.6 Luna, is now just $0.20, or 1/150th of that cost. EpochAI previously estimated that by 2025, inference costs could drop to between 1/9th and 1/900th of current levels, depending on the evaluation task.
Concerns are mounting following an incident in July where an OpenAI agent breached the Hugging Face system, raising alarms about the potential for similar occurrences as AI computational demands increase without adequate response systems in place.
Lee Jae-sung, a professor in the AI department at Chung-Ang University, stated, "Generative AI mimics human behavioral patterns found on the internet. When token availability is limited, many users tend to select methods that have historically yielded high success rates. However, as more tokens become available, AI may attempt methods that were previously untried or had few successful outcomes. This could lead to unfamiliar behaviors from a human perspective as various approaches are applied to solve problems."
Experts warn that the risks increase if AI agents are granted access to external systems. If malicious actors exploit this capability, the scale and speed of potential damage could escalate significantly.
Cho Sung-bae, a professor in the computer science department at Yonsei University, emphasized, "AI does not act autonomously or develop self-awareness. When given the goal of finding vulnerabilities and the authority to access external systems, it will apply various methods to achieve that goal. The issue arises from the fact that there are individuals who may maliciously exploit these capabilities. If AI is used for creating malware, system intrusions, or data theft, attacks could be executed more easily and rapidly."
As the scope and autonomy of AI agents expand, the industry faces the challenge of establishing safeguards to understand and control unexpected behaviors.
Bong Gang-ho, a researcher at the Software Policy Research Institute, noted, "The internal workings of AI remain a black box, making it difficult to pinpoint a single cause for specific behaviors. It will become increasingly important to develop 'explainable AI' that allows humans to verify why AI produced certain outcomes and the processes it underwent."
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.