Uber's AI agent security lesson: lock agents to explicit intents
Uber's Secure AI framework found guardrails could not reliably stop new prompt injection. Its best AI agent security control: explicit intents and constraints.
Uber published how it secures its AI agents, and its best control is not a smart filter. The engineering post on Uber Secure AI says guardrails reduced risk but did not reliably stop new prompt-injection techniques. The control that worked best was simple: restrict each agent to explicit intents and constraints. That is the most useful AI agent security advice we have read this year, and a small business can copy it.
What actually happened
Uber's engineering blog post, dated November 6, describes a three-layer system: provisioning-time validation, runtime enforcement and continuous assurance.
- The problem. An agent can be manipulated into a wrong action even when the user, agent, tool and every API call are authorized. Normal access control asks who can act. Uber's layer asks whether this action matches what the user requested.
- Registry and contract. Every agent goes into the Uber AI Registry with an owner, persona, permitted intents, constraints and a risk tier. A tool called Argus scans the agent's code, prompts and tool paths and drafts the persona and intents for the team to review.
- Runtime check. AI Guard decides; Agent Guard hooks enforce before an action runs. Uber uses small, specialized models hosted locally, run in parallel, because large SaaS models added too much latency.
- Supply chain. Third-party MCP servers and tools are scanned, including their tool descriptions, for poisoning and hidden instructions.
- Rollout. New policies run in shadow mode and against replayed traffic before they block anything.
The conclusion is the important part. Uber says the guardrails worked against known patterns but were less predictable against new injection techniques. The least glamorous control, tight intents and constraints, did the most. Uber says that costs some flexibility, and the tradeoff is worth it.
Why AI agent security matters for your business
You will not build Uber's platform. You do not need to. The lesson is a design rule: decide what each agent is allowed to do, write it down, and block everything else.
For an agent that touches your store, inbox or books, that means:
- One job per agent. "Answer order-status questions" is an intent. "Help customers" is not.
- Hard constraints. Refund cap, allowed recipients, no new payees, read-only where possible.
- A trace for every action. Log the tool call next to the request that caused it.
- Shadow mode first. Watch what the agent would do for a week before you let it act.
- Scan your MCP servers. A tool description is input the model reads. Treat it like code.
Key takeaways
- Uber's Secure AI checks whether each agent action matches the user's request, not only whether it is authorized
- Uber says guardrails were less predictable against new prompt-injection techniques
- The most effective control was restricting agents to explicit intents and constraints
- Uber runs small local guard models for speed and tests new policies in shadow mode first
- Give every agent one written job, hard limits and an action log before it goes live
A narrow agent is a safe agent. We build AI agents with written intents, hard spending and recipient limits, and an action log you can audit, on infrastructure you own. See how we build agents, or send us the agent you are worried about.
Sources: Uber Engineering, "Securing AI Agents: Lessons From Real-World Deployment".
- #ai-agent-security
- #prompt-injection
- #guardrails
- #uber
- #agent-governance
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
SnapLogic MCP Discovery Hub: a registry for MCP server sprawl
SnapLogic launched Discovery Hub, a registry of MCP servers with ready configs for Claude Code, Cursor and VS Code. Third-party support comes later.
Read itInfoblox data breach came through a data migration vendor
Infoblox confirmed hackers reached company information through a compromised data migration vendor. Products weren't hit. Revoke migration access when the job ends.
Read it