OpenAI at Guidewire Connections 2026: AI agent security in insurance is built from multiple layers
At the Guidewire Connections 2026 conference in Las Vegas, OpenAI representative Viyat Bhalodia explained how to protect AI agents in insurance from hidden instructions in documents and errors.
On October 30, 2026, at the Guidewire Connections conference in Las Vegas, Ben Chang, Vice President of Engineering at Guidewire, spoke with Viyat Bhalodia, a member of technical staff at OpenAI. The session was titled "Agent Security 101." Bhalodia said he leads the security direction for Codex and agent products based on Codex and ChatGPT.
The occasion was a specific scenario. An AI agent (a program that autonomously performs the steps of a task) processes an insurance case: it reads a police report, medical records, a repair estimate. What happens if one of the documents contains a hidden instruction that the agent should not follow? Chang and Bhalodia discussed this against the backdrop of insurers moving from pilots to industrial operation of agents.
Previously, such systems more often worked with rigidly defined inputs and outputs. An agent with access to files and the authority to perform operations is a different risk class. Bhalodia said: "You shouldn't rely on just one layer of software protection, because if it fails, you'll be completely vulnerable." According to him, models are trained to resist such attacks, but insurers still need to plan for an attack getting through.
In practice, this means restrictions on file and network access, limits on transaction authority, temporary credentials, automatic verification and escalation to a human when the stakes are high. Chang promised participants an assignment for their organizations.
What remains unknown: which specific models and versions were mentioned, how resistance to attacks was measured, and what implementation timelines were discussed. There is no data on this in the note.