Trend Micro: AI Agent Cyber Risks & LLM Vulnerabilities

Trend Micro and PwC Consulting released a new report highlighting cyber risks in AI agents. They found that certain cyber attacks can succeed across multiple large language models. Here's the thing: The report suggests this is a structural risk in agent-based AI systems. Trend Micro's TrendAI brand ran 2,600 tests using 200 attack prompts. They tested 13 AI models from four different vendors. The tests confirmed that attacks were successful on models from multiple vendors. What's interesting is that the findings indicate stored prompt injection is not tied to a specific model or vendor. Instead, it seems to stem from the architecture of AI agents. In one demonstration, an AI agent was tricked into exfiltrating confidential information, including authentication tokens and other customer data. The bottom line: Current language models have an inherent property where they struggle to distinguish between data and instructions, which impacts the security of AI agents.