Back to Speakers
Divya Ravi -  at Unity

Connect

November 23–27, 2026

Divya Ravi

Unity

Session Details

On-Device Agents and the Lethal Trifecta: What Changes When the Agent Runs on the Phone, Not the Cloud

The "lethal trifecta" is a useful lens for AI agent security: an agent becomes dangerous not from any single capability, but from the combination of three — access to private data, exposure to untrusted content it might follow instructions from, and the ability to send results somewhere outside the user's control. Most discussion of this trifecta assumes a cloud-hosted agent, where the mitigations are familiar: network egress monitoring, sandboxed execution, centralized logging. On-device agents — the kind now shipping via Apple's Foundation Models and Core AI frameworks, running inference and tool calls directly on a user's phone — break most of those assumptions. There's no network chokepoint to monitor if the "exfiltration" is a local call to Messages or Health. There's no centralized log if the whole loop runs on hardware you don't control. And patching a vulnerable agent means an App Store review cycle, not a server-side deploy. This talk walks through how each leg of the trifecta changes shape when the agent moves from cloud to device, using concrete examples from building consent- and permission-aware mobile infrastructure. It covers which cloud-native security assumptions quietly stop being true on-device, and a concrete pattern — treating every on-device tool call as requiring an explicit, auditable permission grant rather than an implicit one — for engineers shipping agentic features into mobile apps today.

Full conference agenda

See every session across the AI Agents Conference November 2026 line-up.

View Agenda