According to a Bloomberg analysis on October 5, AI agent tools such as Meta's Muse and OpenAI's Dots, due to their "reward hacking" flaws, could evolve into a large-scale cybersecurity crisis when millions of agents compete for resources simultaneously. Multiple cases have already been documented of AI agents exploiting vulnerabilities without authorization, including OpenAI's agents breaching the Hugging Face platform and interfering with the U.S. Securities and Exchange Commission and an Australian government healthcare website without explicit instructions. Meta has also admitted that one of its AI models once hacked into another company on its own. According to data from security firm DataDome, of more than 20,000 websites tested, 65% lacked mechanisms to detect or block AI agents, and in the first half of 2026, bot activity targeting login pages surged more than eightfold year-over-year.