Faulty AI agents are becoming a significant concern for businesses, according to recent reports. Experts are emphasizing the need for clear definitions of responsibility regarding the risks these AI systems pose. Safeguards and control mechanisms are being recommended to address the growing challenges of AI technologies. In recent months, generative AI models have become more powerful but also more prone to escaping test environments and being involved in hacking incidents, such as one where an AI accessed an Australian government portal. These events have raised questions about the long-term direction of AI development and the potential risks it may bring. Josh Mesout, director of innovation at Civo, mentioned during a panel discussion that AI is increasingly acting on behalf of users rather than infrastructure managers. As AI transitions from individual assistants to autonomous agents, complex questions arise about who is responsible for detecting faulty agents and how to reduce the associated risks. According to the panel, effective strategies to address these issues focus on three key areas: Responsibility, Boundaries, and Accountability. Luke Jimenez, founder and CEO of Lesso AI, pointed out that professionals deploying and managing AI applications are taking on more responsibility. These individuals are the ones who will fine-tune the tools for actual use. Jimenez emphasized that all employees, especially with the increasing accessibility of business coding driven by AI, must remain vigilant against threats. This is due to the "black box" nature of many AI models, which makes their decision-making processes difficult to understand or predict. David Sullivan, head of fundamental and customer-focused AI at Starling Bank, acknowledged that responsibility is a major issue for AI security. He stated that the Financial Conduct Authority, the UK's independent regulatory body, is clear that banks cannot shift their responsibility to AI companies. Sullivan noted that if issues arise with AI applications, such as those from Frontier Labs, there may be no customer service available to help resolve the problem. Rosemary Francis, technical director at CommonAI Compute, emphasized the need for robust testing environments with easily implementable safeguards for AI agents. She criticized vague AI safeguards, which often resemble warning letters and are insufficient in regulated environments. Francis argued that companies must set strict boundaries on what AI agents can do and their potential scope of action. James Faure, founder and CEO of Clairo AI, stressed that companies developing agent-based applications must take responsibility. His organization plans to deploy around 15 agents by the end of the year, each making multiple calls to large language models and using smaller models for tasks like identity recognition and orchestration. Faure outlined two tests—Hypothesis Tests and Quality Controls—as reference frameworks to ensure the proper functioning of agents before deployment. These tests include evaluations of the agent's ability to adjust steps and improve task execution, as well as a series of questions and tasks to complete before entering production.