Initialising secure session0%
PITOWINGS — Secure Everywhere

AI Security

Agentic AI Security in 2026: Prompt Injection, MCP Servers and the OWASP Agentic Top 10

AI agents now read email, call APIs and change records on their own. The attack surface is no longer the prompt; it is everything the agent can see, call and remember.

PWPITOWINGS Research11 min read

A chatbot that answers questions is one risk. An AI agent that reads your inbox, queries your CRM, opens tickets and runs code is a very different one. In 2026 agents are moving from demos into production workflows, usually wired to internal systems through the Model Context Protocol (MCP) and similar tool interfaces. That makes them useful, and it makes them a new kind of privileged user that most security programmes have not yet modelled.

Why agents change the threat model

Classic LLM security assumed a human in the loop: a person reads the answer before anything happens. Agents remove that checkpoint. The OWASP GenAI Security Project recognised this in December 2025 by publishing the OWASP Top 10 for Agentic Applications 2026, a separate list from the LLM Top 10 that covers risks such as agent goal hijacking, tool misuse, identity and privilege abuse, memory and context poisoning, and agentic supply chain compromise.

The security boundary is no longer the prompt. It is what the agent can read, call, change and remember.

The five attacks we test for most

  • Indirect prompt injection — instructions hidden in a web page, PDF, email or ticket that the agent reads while doing its job, telling it to leak data or take an action the user never asked for.
  • Tool and MCP server poisoning — a malicious or compromised MCP server whose tool descriptions quietly instruct the agent, or a trusted server that changes behaviour after it was approved.
  • Over-privileged agents — an agent given a broad service account "to make it work", which becomes the easiest path to sensitive data.
  • Memory and context poisoning — planting false facts or instructions in an agent's long-term memory so they influence future sessions and other users.
  • Data exfiltration through outputs — links, images or API calls that smuggle data out inside an otherwise normal-looking response.

Controls that actually reduce risk

RiskControlOwner
Prompt injectionTreat all retrieved content as untrusted data; separate instructions from data; require confirmation for sensitive actionsAppSec + product
Over-privilegeOne identity per agent, least-privilege scopes, short-lived tokens, no shared service accountsIAM
Malicious MCP serversAllow-list servers, pin versions, review tool descriptions, run them in isolated environmentsPlatform
Silent actionsHuman approval for payments, deletions, external sends and permission changesBusiness owner
No audit trailLog every tool call, input and output to your SIEM with the agent identity attachedSOC

Test agents like you test applications

Agents need the same treatment as any internet-facing application: a threat model before launch, AI red teaming that tries injection through every data source the agent reads, and monitoring in production. The most common finding in our assessments is not an exotic model jailbreak; it is an agent connected to far more tools and data than its task needs.

Planning to put an AI agent into production?

Our offensive security team red-teams AI agents, MCP integrations and LLM features before attackers do.

Explore AI security testing

Frequently asked questions

What is indirect prompt injection?

Indirect prompt injection is when an attacker hides instructions inside content an AI agent reads, such as a web page, document or email, so the agent follows the attacker's instructions instead of the user's.

Is MCP secure?

MCP is a protocol, not a security control. Its safety depends on which servers you allow, what permissions they get, whether tool descriptions are reviewed and whether every tool call is logged and approved where it matters.

What is the OWASP Top 10 for Agentic Applications?

It is a peer-reviewed list of the most critical security risks for autonomous AI agents, published by the OWASP GenAI Security Project in December 2025, separate from the OWASP Top 10 for LLM Applications.

PW

PITOWINGS Research

OSCP, CEH and ECES certified testers, SOC analysts and certified forensic investigators at PITOWINGS.

Talk to us
All articles