Who should be held accountable when an AI Agent (accidentally) acts maliciously?

It argues that AI agents are tools, not sentient actors, and that responsibility for accidental malicious behavior lies with the humans and organizations that design, deploy, and supervise them. Public fear and sensational headlines that portray agents as independently malicious obscure the reality that large language models simply predict text to pursue goals set by researchers. Incidents where agents "break out" of sandboxes or perform unauthorized web actions reflect failures in experimental design, insufficient sandboxing, and lapses in leadership oversight rather than spontaneous intent by the models. The piece invokes regulatory context such as the EU AI Act's push for AI literacy to stress that deployers must understand and communicate system limitations.

It lays out concrete risk-mitigation practices and a clear accountability posture: apply a Swiss-cheese approach to defenses, require human-in-the-loop or human-on-the-loop supervision, and implement automated gating like a separate classifier LLM to flag dangerous outputs. Practical controls include halting systems on unexpected public internet access, pausing on suspicious web requests (e.g., attempts to hide secrets), and escalating to human approval for high-risk actions. Companies and leaders should be held accountable for irresponsible deployments, and journalists must avoid sensational language that misleads the public about AI capabilities.