Enforce Guardrails for AI Agents

Enforce Guardrails for AI Agents

Agentic AI is moving from pilots into production across industries quickly. In finance, for example, agentic AI is being used for customer service, fraud triage, and loan application processing. In pharmaceuticals, it is used for drug discovery. In insurance, for processing claims submissions. In software and IT, for coding. The list goes on and continues to grow.

Anjuna AI

Architects of Trust: Fireside Chat with Industry Leaders

Confidential Computing Summit 2024

Mark Bower, VP of Product at Anjuna, sits down with leaders from the financial services, blockchain, and security sectors to discuss their adoption and utilization of Confidential Computing technology.

Left click drag to rotate. Two-finger drag / mouse wheel to zoom. Right click drag / Command (or Control) drag to pan.

What these uses have in common is that agents hold credentials, call tools, take actions, and consume paid resources at machine speed, faster than any human reviewer can follow.

What these uses have in common is that agents hold credentials, call tools, take actions, and consume paid resources at machine speed, faster than any human reviewer can follow. Agents can become vectors for vulnerabilities and increase the attack surface; for example, prompt injection can steer an agent into misusing its access or leaking data. They can also cause non-compliance and runaway spend.

That is why organizations are looking for stronger controls to manage security, compliance, and financial risk, as they all originate from the same place: what agents can access, do, and spend. The security and compliance controls should autonomously prevent misuse in real-time, because detecting the breach or incompliance after the fact is note nough. Worse, not having a trail of what the agents did will prevent the team to plug the security hole.
‍

The financial controls that limit AI spend work with the security and compliance controls for a system of checks and balances. By limiting the token spend of a compromised agent or one that is creatively trying to achieve an intended objective, the damage can be capped. As well, enterprises are starting to demonstrate financial responsibility on AI spend, including Accenture, which centrally funds Claude Enterprise use up to $29 per employee per month as of October 1. Bloomberg reported in June that Uber had capped employee use of AI tools like Claude Code to manage costs. Other examples include Microsoft and Atlassian.

‍
However, the governance controls still harbors a major vulnerability. The concern is how to protect the guardrail system itself so that it cannot be hijacked or tampered with. Anjuna Overwatch runs within a hardware-rooted trust boundary of Anjuna Seaglass and acts as a tamper-resistant control plane. Because Anjuna Overwatch runs in a confidential computing environment, policies can be changed only through an authorized, auditable process, and not even privileged administrators can override them. Agent governance shifts from something you monitor to something you can prove and enforce automatically.

Anjuna Overwatch manages the risks of autonomous agents in the following ways:

  • Prevent unauthorized access: Anjuna Overwatch inspects every LLM message and tool call in real time inside your own confidential computing environment, so no one, including Anjuna, can see the content. Because it is self-managed, Anjuna is not in the chain of trust and there are no round trips to Anjuna, which keeps performance high.
  • Prevent data misuse: Anjuna Overwatch enforces least-agency access to data and applications, so each agent reaches only the data its task requires
  • Avoid credential abuse: Anjuna Overwatch holds credentials on the agent’s behalf, so the agent never touches them. A compromised or misbehaving agent has no credential to leak or misuse.
  • Control financial risks: Anjuna Overwatch tracks token spend, tool-call volume, and compute consumption per agent, per task, and per business unit. It can adjust or throttle an agent before it exceeds its budget, and automatic circuit breakers stop runaway loops, such as an agent stuck in a retry loop that would otherwise burn through its budget overnight. Because Anjuna Overwatch understands context and history, it can also match the right model to each task, reaching the goal at the lowest cost.
  • Support regulatory compliance: Anjuna Overwatch maps each action to policy and compliance policy as it runs. The controls align with regulatory frameworks including NIST 800-53, the NIST AI RMF, FINRA rules, and the EU AI Act, with evidence produced in real-time. IT, security, and finance share a single tamper-resistant record of what each agent accessed, what it spent, and under whose authorization.

Left click drag to rotate. Two-finger drag / mouse wheel to zoom. Right click drag / Command (or Control) drag to pan.

Enforce Guardrails for AI Agents

Case Studies

All Case Studies
No current matching case studies. Check back soon...

Customer Success

All Case Studies
No items found.

Ready to get started?

Learn more about Anjuna, register for a live demo, or contact sales