Skip to content

Meet DeepintShield at Bengaluru Tech Summit 2026, 17 to 19 November, BIEC Hall 1, booth H1A106. Book 20 minutes

Start For Free

Control what AI agents can do

Give each agent a verified identity and authorise every tool action against exact relationships, with Allow, Deny or Require approval.

The reliable answer to how to control what AI agents can do is to authorise each action, not each agent. DeepintShield puts a single decision point in front of every governed tool call, checks the agent’s verified identity and its exact relationships, and returns Allow, Deny or Require approval before anything runs.

The challenge

Agents turn model output into actions. That is their value, and it changes the security question from “what did the model say?” to “what is this agent allowed to do?”:

  • Broad credentials. Agents often run with a service account that can do far more than the task needs.
  • Text becomes action. A prompt injection in a document can steer an agent towards a tool call nobody intended.
  • Tool names hide risk. One tool may read a record or delete it, depending on its arguments.
  • Unknown agents. Teams build agents faster than security can list them, so some run with no owner at all.

The approach

Give every agent an identity it has to prove. Grant permissions on exact objects, tools and actions, and deny everything else. Hold consequential actions for a person. Fail closed when the decision cannot be made. Record each decision before the call proceeds. Start in advisory mode to see what would be denied, then enforce. A practical order of work: list the agents already calling models, name an accountable owner for each, write down the actions each one needs, and grant only those. The pattern is explained in how PEP and PDP protect tool calls.

How to control what AI agents can do with DeepintShield

  1. Give each agent a verified identity. Register agents, or auto-register a verified class of agents with enrolment policies. Agents prove a provider-issued identity through Microsoft Entra Agent ID or generic OIDC (for example Okta, Auth0 or Keycloak), with a one-hour token lifetime ceiling by default. Trust based only on a Virtual Key is meant for local development. CI tokens used for registration need an expiry and cover one workspace. See agent workload identity.
  2. Grant exact relationships. OpenFGA checks the agent’s permission on the object, can_use on the tool and can_execute on the action, plus delegation legs when an agent acts for a user. Every leg must pass, and errors fail closed. Discovered actions count as write and high risk until an operator classifies them; an agent never classifies its own tools. Task mandates can narrow access for one job but never grant it. See relationship authorisation.
  3. Enforce in code and roll out safely. With Python SDK 3.2.3, wrap a function with @shield.agentic.tool("ledger.post", action="create_entry", action_class="write", recovery_cost="high"), or let automatic enforcement cover LangGraph, LangChain, LlamaIndex, AutoGen/AG2, PydanticAI, OpenAI Agents SDK, AWS Strands, Google ADK and Temporal. Use Advisory mode only for migration, then switch to Enforcing.

Agentic Security is available on the Business plan and above. SDK instrumentation is cooperative defence in depth, so also remove direct credentials that would let an agent reach a service around the gateway.

Test the boundary with three calls: an allowed read, a denied write, and a call while the authorisation service is unavailable. The last one should fail closed with mcp_tool_authorization_unavailable, not run.

What you can show afterwards

  • Every agent, its accountable owner and its status, such as live, pending or quarantined.
  • Each decision with its outcome, recorded in an append-only, hash-chained ledger before the call ran.
  • Decisions classified against the OWASP Agentic Top 10, with crosswalks to NIST AI RMF, ISO/IEC 42001, the EU AI Act and MITRE ATLAS.
  • Runs, Decisions and Trends views, with CSV export for reviewers.
  • Agents quarantined after a code change added new findings, and who approved them again.
Start For Free