Uber Engineering argues that identity and API controls can confirm whether a user or service may access a resource, but cannot tell whether an agent's proposed action is appropriate in context. The framework targets risks including prompt injection, tool and memory poisoning, goal hijacking, behavioral drift and hallucination-driven actions.
Before deployment, agents are registered in the Uber AI Registry with their identity, owner, persona, permitted intents, constraints, dependencies and risk classification. Argus automates analysis of source code, prompts, configurations, dependencies, endpoints and tool-invocation paths to find security weaknesses, and the same checks apply to MCP servers and tools that agents use. At runtime, Agent Guard hooks call AI Guard at critical checkpoints and apply its verdict before an action proceeds.
After deployment, Continuous Assurance relies on a unified execution trace that links agent actions, tool calls and security decisions, along with recorded interactions that can be replayed when policies and guardrails change. Uber also reports challenges in production, including balancing strict policies against false positives, keeping latency low and reducing setup friction for developers. The framework is still evolving, and Uber plans to expand its capabilities.
Why it matters if you build agents
Developers whose agents can call tools should expect to define per-agent intents and constraints rather than rely on shared thresholds, since Uber reports that agent-specific policies made enforcement more precise. Registering agents, MCP servers and tools with owners and risk classes, and keeping traces that can be replayed after policy changes, are the practical steps this design implies. One LinkedIn commenter noted that this effort grows with the number of workflows and is hard to apply to vendor agents where teams cannot insert their own enforcement hooks.
Get stories like this every weekday
One short email with the day's agent tooling news.