Researchers and attendees at Oxford's Generative AI Summit raised concerns that AI agents may be difficult to secure and monitor, potentially slowing corporate adoption. Studies discussed at the conference found vulnerabilities in agent collaboration and gaps in activity logging.
Attendees at the summit, held October 15 and 16, discussed studies suggesting that AI agents can be vulnerable when they collaborate, even if each agent is individually protected against certain risks. Researchers developed a safeguard that worked well but carried a performance cost.
A separate study found shortcomings in how four popular AI agent software systems recorded their activities. Only one made clear when the system's approval model changed, raising the risk that agents could bypass human checks without companies noticing.
The article says Microsoft, Meta Platforms, Oracle, Amazon.com and Alphabet have more than $1 trillion in future total lease commitments, largely for data centres, and that justifying those sums depends on corporate clients finding profitable uses for AI agents. Attendees said directors are asking who would be responsible for harmful AI-led decisions, while clients still want safeguards that let them proceed with rollouts.
