Facilitated the Internet Resilience Forum at UCLA in Los Angeles. One of the key lessons: artificial intelligence is fast becoming a new layer of systemic risk. As the Internet shifts toward agents that communicate, negotiate, and act on behalf of users and organisations, failures will happen more frequently. We tend to anthropomorphise these systems, often trusting them beyond what is warranted, amplifying the risk of misuse or misunderstanding.
Specific vulnerabilities are emerging: hidden prompts that subtly subvert agent behaviour; agent-to-agent “drift” that distorts original intent; and the growing difficulty of tracing errors across automated chains.
Managing this next layer will require shared architectural standards, including Agent-to-Agent (A2A) protocols and the Model Context Protocol (MCP), to ensure predictable behaviour and enable verifiable interactions.
📸 With Vint Cerf. In the gallery, a visit to UCLA Boelter Hall 3420, where the first message was sent over the Arpanet. With Leonard Kleinrock, Nick McKeown and David Cross.