A new paper on arXiv, titled “Towards Safe Autonomy for Computer Use,” highlights a critical gap in the safety of LLM-based agents. According to the abstract, the safety of these agents currently rests on two mechanisms: alignment training and safety classifiers. The paper argues that neither of these can rule out the possibility that a sophisticated indirect prompt injection (IPI) gets through and causes significant harm.
Indirect prompt injection is a known attack vector in which malicious instructions are hidden in content that an agent retrieves or processes, rather than in direct user prompts. The authors suggest that current defenses are insufficient for the high-stakes setting of autonomous computer use, where an agent may act on untrusted web content, emails, or files.
The abstract is brief and cuts off after noting recent work, so the full proposed solution is not visible in the announcement. Still, the framing makes clear that the paper is motivated by the limits of existing safety measures and the need for more robust approaches before agents can be safely deployed for computer-use tasks.