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

Platform · Agentic Security

Agentic AI security: a verdict before every tool call

Agentic AI security controls what autonomous agents may do, not just what they say. DeepintShield checks every tool call an agent makes against identity, relationships and policy, then returns Allow, Deny or Require approval before anything runs. Each decision is written to a hash-chained ledger, so you can prove what happened.

  • Allow, Deny or Require approval
  • Record-level authorization with OpenFGA
  • Hash-chained decision ledger
Agentic AI security illustration: an AI agent beside a shield with allow, deny and approval icons

Why chatbot guardrails do not protect AI agents

Agents read untrusted content, hold credentials and act on real systems. Content filters inspect text; they do not decide whether an action should happen.

Action-targeted injection

Hidden instructions in a web page, file or tool result push the agent to call the wrong tool with the wrong arguments.

Over-privileged agents

Agents inherit broad service credentials, so nothing limits them to the records and actions their task needs.

Tools that change after approval

An MCP server can change a tool’s description or schema after review, and the agent keeps calling it.

How agentic AI security works

One decision point sits in front of every governed tool call, whether it comes from the SDK, the MCP gateway, Agent Mode or Code Mode. It is part of the wider DeepintShield platform.

Identify the agent

A Microsoft Entra Agent ID or OIDC token identifies the agent, not a name it chose itself.

Intercept the call

The SDK or the gateway holds the tool call before the function body runs.

Check every leg

Key binding, action class, OpenFGA relationships and optional OPA or Cedar policy must all allow it.

Record the verdict

The decision is appended to a hash-chained ledger, then the call runs, stops or waits.

Three verdicts for every agent action

Adding an agent, a tool or an MCP server never grants access by itself. Each governed call needs an exact Allow.

Allow

The call runs

Identity, relationships and policy all agree, so the tool runs and the decision stays on record.

Deny

The call stops

One failed check denies the call. Approvals, auto-execute lists and OAuth grants cannot turn a Deny into execution.

Require approval

A person decides

Write-like or unclassified actions wait for a designated approver. The requester and the agent’s owner cannot approve.

Pending approvals expire after five minutes by default, and you can extend that to 24 hours.

Agentic security features

Agent identity

Agents prove identity with Microsoft Entra Agent ID or any OIDC issuer, with a one-hour default token-lifetime ceiling.

Policy-based enrolment

Enrolment policies register verified agents with an owner, risk tier and namespace. Every call is still authorized.

Record-level authorization

OpenFGA relationship checks confirm the agent may use this tool, run this action and touch this record.

Pinned tool definitions

Approval binds a tool’s name, description and argument schema. Any change denies calls until someone re-approves it.

Code review and quarantine

Agent code gets a mandatory deterministic scan. Incomplete coverage fails closed, and new findings quarantine the agent.

Advisory, then enforcing

Start in Advisory (shadow) mode to see verdicts, then switch to Enforcing. Advisory never skips required approvals.

Agentic Security, agent enrolment and OpenFGA authorization are available on the Business plan and above, where they are on by default. Developer and Team workflows run without Agentic Security. Compare plans and pricing.

One decision point for agents and MCP tools

A policy enforcement point (PEP) holds the call; a policy decision point (PDP) evaluates it. The same check covers SDK-governed tools, JSON-RPC tools/call on the MCP gateway, the direct tool-execution endpoint, Agent Mode and every nested Code Mode call.

  • Every leg must pass, and errors fail closed
  • New actions count as write, high risk until classified
  • An agent never classifies its own tools
Agentic AI security decision flow: the enforcement point holds an agent tool call while the decision point checks identity, OpenFGA relationships and context policy; the result is Allow, Deny or Require approval, recorded in the ledger(opens the full-size image in a new tab)
Decision flow for one agent tool call, from interception to the hash-chained audit log.

Pinned tool definitions and an AI bill of materials

When you approve a tool, DeepintShield pins a digest of its name, description and argument schema. If the server later advertises a different definition, the console shows both digests and calls stay denied until an operator chooses Approve new definition.

Export AIBOM gives you a JSON inventory of every governed tool contract. It is signed when a signing key is configured; otherwise it is clearly marked unsigned.

Agentic AI security tool pinning: the approved tool definition’s digest is compared with the live one on every call; a match goes ahead, drift is denied until re-approved(opens the full-size image in a new tab)
A changed tool definition is denied until it is re-approved.

Govern a tool with one decorator

The DeepintShield Python SDK 3.2.3 (Python 3.10 or newer) adds a pre-call check to your own functions. Automatic enforcement covers LangGraph, LangChain, LlamaIndex, AutoGen/AG2, PydanticAI, OpenAI Agents SDK, AWS Strands, Google ADK and Temporal.

  • Preflight and authorization run before the function body
  • Set DEEPINTSHIELD_AGENT_NAME: there is no default
@shield.agentic.tool(
    "ledger.post",
    action="create_entry",
    action_class="write",
    recovery_cost="high",
)
def post_entry(row: dict) -> dict:
    return ledger.insert(row)

Mapped to agentic and AI risk frameworks

Mapped to

  • OWASP Agentic Top 10
  • OWASP LLM Top 10
  • NIST AI RMF
  • ISO/IEC 42001
  • EU AI Act
  • MITRE ATLAS

Agent decisions are tagged against the OWASP Agentic Top 10, with crosswalks to NIST AI RMF, ISO/IEC 42001, the EU AI Act and MITRE ATLAS. Mapping supports your own assessment; it does not replace it.

Further reading on agent security

Agentic AI security questions

What is agentic AI security?

Agentic AI security is the set of controls that decide what an AI agent may do with tools, data and other systems. It goes beyond content filtering: each action is checked against identity, relationships and policy before it runs, and the decision is recorded. See how DeepintShield works.

What is a policy enforcement point for AI agents?

A policy enforcement point (PEP) is the component that holds a tool call until a policy decision point (PDP) returns a verdict. In DeepintShield, the SDK and the gateway act as enforcement points, and one decision point serves SDK tools, MCP calls, Agent Mode and Code Mode.

How does human approval work?

A Require approval verdict holds the call until a designated approver accepts or rejects it. The requester and the agent's accountable owner cannot approve, an approval cannot override a hard denial, and the approval is tied to a digest of the call's arguments.

Which agent frameworks are supported?

The Python SDK 3.2.3 governs tool calls in LangGraph, LangChain, LlamaIndex, AutoGen/AG2, PydanticAI, OpenAI Agents SDK, AWS Strands, Google ADK and Temporal. CrewAI is not supported. Details are in the Agentic Security docs.

Which plan includes Agentic Security?

Agentic Security is available on the Business plan and above, where it is on by default. On the Developer and Team plans, agent workflows run without Agentic Security checks. See the plan comparison for details.

Give every agent action a verdict

Start free on the Developer plan, or book a demo to see Agentic Security on Cloud Business.

Qualified organisations can ask for a complimentary 30-day Cloud Pilot on Cloud Business.

Start For Free