Tools and actions
Agents call APIs, run code and write to systems. Every tool an agent can use is something an attacker can try to use through it.
AI agents do not just answer. They plan, call tools and act with real permissions. Here is what agentic AI security covers, which risks are new and how to control them.
Agentic AI security is the practice of protecting AI systems that act on their own: agents that plan steps, call tools, keep memory and hand work to other agents.
It covers the agent itself, the tools and data it can reach, the permissions it runs with and what it does at runtime. Because an agent turns model output into action, agentic AI security adds controls on behavior to the controls that protect models and data.
A chatbot that gets something wrong produces a bad answer. An agent that gets something wrong sends the email, runs the query or changes the record. Four things separate agents from the AI applications most security programs were built for.
Agents call APIs, run code and write to systems. Every tool an agent can use is something an attacker can try to use through it.
Agents chain many steps without a person approving each one, so a single bad instruction can travel a long way before anyone looks.
Agents carry context from one step and one session to the next. Content planted in memory can steer decisions long after it arrived.
Agents hand work to other agents and reach tools through MCP, an open-source standard for connecting AI applications to external systems.
Read the MCP postAdoption has moved faster than control. Independent research shows how often agents do things nobody intended, and how few organizations have rules for them.
In a 2025 survey of 353 enterprise IT and security professionals, 82% of organizations said they already use AI agents.
SailPoint, 202580% of companies say their AI agents have taken unintended actions, and 39% say an agent accessed systems or resources it was not authorized to use.
SailPoint, 202523% of companies say their AI agents were tricked into revealing access credentials.
SailPoint, 202596% of technology professionals consider AI agents a growing risk, yet only 44% of organizations have policies in place to secure them.
SailPoint, 202513% of organizations reported breaches of AI models or applications. Of those, 97% did not have AI access controls in place.
IBM, 2025In May 2026 the NSA published security design guidance for MCP, naming dynamic tool invocation and implicit trust between agents among the risks.
NSA, 2026In December 2025 OWASP published a Top 10 written for agents, separate from its list for LLM applications. It gives security reviews a shared checklist for agents. See how Cranium AI maps to each risk, including where it stops.
| ID | Risk | What it means |
|---|---|---|
| ASI01 | Agent Goal Hijack | An attacker redirects what the agent is trying to achieve. |
| ASI02 | Tool Misuse & Exploitation | Legitimate tools get used in unintended, harmful ways. |
| ASI03 | Identity & Privilege Abuse | Agents inherit, escalate or misuse credentials and permissions. |
| ASI04 | Agentic Supply Chain Vulnerabilities | Compromised frameworks, tools, MCP servers or agent configs. |
| ASI05 | Unexpected Code Execution | Agents generate or run code an attacker steered. |
| ASI06 | Memory & Context Poisoning | Malicious content persists in agent memory and steers later steps. |
| ASI07 | Insecure Inter-Agent Communication | Messages between agents get spoofed, tampered or abused. |
| ASI08 | Cascading Failures | One bad step propagates across agents and systems. |
| ASI09 | Human-Agent Trust Exploitation | Agents manipulate the people who supervise them. |
| ASI10 | Rogue Agents | Agents drift from intended behavior or act outside scope. |
Agentic AI security builds on LLM application security. The threats to the model are still there. What grows is the reach of a mistake.
| LLM application security | Agentic AI security | |
|---|---|---|
| What you protect | Prompts, responses and the data behind them | All of that, plus tools, memory, permissions and other agents |
| What goes wrong | A harmful or leaked answer | A harmful action taken in a real system |
| How far it spreads | One conversation | Every system the agent can reach, and the agents after it |
| Human in the loop | A person reads each answer | Many steps run before a person looks |
| Where to look | The prompt and the response | Every event: user, model, tool, memory and agent to agent |
| Reference list | OWASP Top 10 for LLM Applications (2025) | OWASP Top 10 for Agentic Applications (2026) |
No single control covers an agent. Cranium AI secures agents with five moves, run continuously, because agents change every time a tool, prompt or model changes.
Inventory every agent, the tools it can call, the MCP servers it connects to and the models behind it. You cannot scope an agent you have not found.
Watch what agents do in production: sessions, tool calls and handoffs. Out-of-pattern behavior shows up here first.
Set policy for each agent event type, with block, modify or pass per signal. Scope tools and permissions to the task in your identity stack.
Red-team agents before release, then enforce policy on live traffic so a hijacked instruction is caught while it runs.
Keep current evidence of which agents you run, what they can reach and how they are tested, ready for auditors and customers.
Each of these goes further into one part of the problem.
What the NSA’s MCP design considerations mean for the agents you already run.
Read the postHow to govern the credentials agents use, and why they are not ordinary service accounts.
Read the postWhere existing application security stops short once AI starts taking actions.
Read the postA practical way to get the return from agents without taking on unbounded risk.
Read the postCranium AgentSensor maps agents, the tools they invoke and the agents they reach, and inventories the frameworks, tools and MCP servers behind them.
See DiscoverSessions, timelines and sequence diagrams show each agent decision and tool call, so you can see where in a chain a risk was introduced.
See ObserveEach agent event type, from user and model to tool, memory and agent to agent, is evaluated under its own policy, with block, modify or pass per signal.
See SecureA risk-by-risk map to the OWASP agentic Top 10 shows what Cranium AI covers, how well and where it stops.
See the OWASP mapStraight answers, kept current. Last updated .
Ask us anything →Agentic AI security is the practice of protecting AI systems that act on their own: agents that plan steps, call tools, keep memory and hand work to other agents. It covers the agent, the tools and data it can reach, the permissions it runs with and its behavior at runtime.
LLM security protects prompts, responses and the data behind them. Agentic AI security covers all of that plus what the agent can do: its tools, memory, permissions and handoffs to other agents. The failure is no longer only a bad answer. It is an action taken in a real system.
The OWASP Top 10 for Agentic Applications, published in December 2025, lists them: agent goal hijack, tool misuse, identity and privilege abuse, agentic supply chain vulnerabilities, unexpected code execution, memory and context poisoning, insecure inter-agent communication, cascading failures, human-agent trust exploitation and rogue agents.
MCP, the Model Context Protocol, is an open-source standard for connecting AI applications to external systems. It gives agents a common way to reach tools and data, which also gives attackers a common path. In May 2026 the NSA published design guidance for MCP that calls for validating parameters, sandboxing tool execution, and logging and monitoring.
Excessive agency is an OWASP risk for LLM applications: a system can take damaging actions in response to unexpected, ambiguous or manipulated model output. OWASP names three root causes: excessive functionality, excessive permissions and excessive autonomy. The fix is to give each agent only the tools, access and independence its task needs.
Start with an inventory of every agent, its tools and its MCP servers. Watch agent behavior in production, set policy for what each agent may do, scope tools and permissions to the task, red-team agents before release and enforce policy on live traffic. Then keep the evidence current.
Cranium ONE inventories agents, tools and MCP servers, shows every agent decision and tool call, and evaluates each agent event against policy at runtime, through the AI Trust Loop: Discover, Observe, Govern, Secure and Prove. Agent identity and privilege management sit outside Cranium AI today, so pair it with your identity platform.
See how Cranium AI discovers, observes, governs, secures and proves every model, agent and vendor in your enterprise.
Cookie Policy · Privacy Policy
Global Privacy Control detected. Opt-out request honored: marketing tags stay off in this browser.