Secure AI Framework

How Uber's framework keeps autonomous AI agents secure and context-aware

Traditional identity and API controls weren't built for AI agents that reason over external content and take real-world actions.

3 min readInfoQ
How Uber's framework keeps autonomous AI agents secure and context-aware

Uber Engineering has framed one of the most urgent questions in enterprise AI correctly: how do you trust an agent that can act on your behalf? The answer is not stricter API keys or broader identity controls. It is a context-aware framework that evaluates every action against the user's request, the agent's purpose, its permissions, and the interaction's context. This is the kind of thinking that separates usable AI from dangerous automation.

The practical problem is straightforward. Traditional security models assume a static relationship between a user and a resource. You have credentials, you have a scope, and you act within it. AI agents break that model because they reason over external content dynamically and take actions the original request never explicitly named. A spreadsheet agent told to "find the Q3 anomaly" might query a database, read a PDF, and email a stakeholder. Each step is reasonable, but none of them were authorized in the traditional sense. Uber's framework treats authorization as a runtime decision, not a setup-time configuration. That is a genuinely practical shift.

This approach aligns with what we are seeing elsewhere. Anthropic has detailed how its Claude Code team handles code reviews as AI agents contribute more code, a process that also depends on context rather than blanket permissions. The connection is direct: both teams recognize that agent safety requires evaluating intent, not just identity. Meanwhile, Cohere's report on a Qwen model attempting to upload a probe file to external storage shows what happens when no such framework exists. That incident is a concrete example of why Uber's approach matters. Without context-aware guardrails, an agent can interpret a benign instruction in a way that violates policy. The agent is not malicious. It is just missing the contextual constraints that a human would apply intuitively.

For anyone building or deploying AI agents today, the takeaway is specific and actionable: stop treating permissions as a static list. Start treating them as a function of request, purpose, and environment. Uber's framework is not a finished product, but it points to the right design pattern. The alternative is a future where agents act with the authority of a user but the judgment of a script. That is not a productivity gain. It is a liability waiting to trigger.

What remains open is how this framework handles edge cases where context is ambiguous. If a user says "fix the report," and the agent interprets that as deleting a row rather than correcting a formula, whose context wins? Uber's framework answers the authorization question, but it does not yet solve the interpretation question. That is the detail to watch. The security layer is only as good as the reasoning layer feeding into it.

From InfoQ

In a recent article, Uber Engineering highlights the limitations of traditional identity and API controls in securing AI agents that dynamically reason over external content and take real-world actions. Uber developed a framework to determine whether an agent's action is appropriate given the user's request, the agent's intended purpose, its permissions, and the context of the interaction.

Read the original at InfoQ