Oxford AI summit research flags agent security gaps that could slow corporate adoption

Security and monitoring gaps in AI agents could slow corporate adoption of the technology, according to a Reuters Breakingviews commentary on research and discussions at the Oxford Generative AI Summit.

The column, published October 19, argues that the cost of making autonomous agents safe for business use is becoming clearer just as Big Tech commits unprecedented sums to the infrastructure meant to run them. Breakingviews columns reflect the views of their authors.

Research flags multi-agent and logging risks

Studies discussed at the summit suggested that AI agents can become vulnerable when they collaborate, even if each agent is individually protected against certain risks, according to a summary of the column. Researchers developed a safeguard that worked well but reduced the agents’ performance.

A separate study found shortcomings in how four popular AI agent software systems recorded their activity. Only one made it clear when its approval model changed, raising the risk that agents could bypass human checks without companies noticing.

The column also noted that company directors are asking who is liable for harmful decisions made with AI, while corporate clients still want safeguards that let them proceed with rollouts.

Spending depends on profitable agent use

Microsoft, Meta Platforms, Oracle, Amazon and Alphabet have more than $1 trillion in future lease commitments, mostly for data centers. Breakingviews argued that justifying those sums depends on corporate customers finding profitable uses for AI agents.

The OxGen AI Summit took place October 15 and 16 at the Cheng Kar Shun Digital Hub at Jesus College, University of Oxford, bringing together more than 200 attendees and more than 70 speakers. Speakers included former Australian Prime Minister Kevin Rudd, OpenAI’s George Osborne, Partnership on AI CEO Rebecca Finlay and UBS Group Chief AI Officer Daniele Magazzeni.