What are the security risks of AI agents?

AI agents act, not just answer. The main risks are hijacked goals, excessive permissions and untrusted tools such as MCP servers, and what to check first.

Jared Henderson avatarByJared Henderson
13 min read

AI agents carry more risk than chatbots because they act: they call tools, use credentials and change real systems. The main risks are agents being hijacked by instructions hidden in what they read, holding more access than their task needs, and connecting to untrusted tools, including through MCP. An agent can do as much damage as its access allows.

The NCSC describes agentic AI as systems that can access data sources, remember context, make decisions, use tools and take actions in pursuit of a goal, without continuous human intervention (NCSC, retrieved 3 November 2026). That autonomy is the point, and it is also the problem.

Executive Summary & Key Takeaways

  • The OWASP Top 10 for Agentic Applications, published in December 2025, puts agent goal hijack, tool misuse and identity and privilege abuse at the top of its list.
  • The NCSC and five partner agencies advise organisations to start small, use agents only for low-risk tasks and never grant broad or unrestricted access to sensitive data or critical systems.
  • Permissions are the control that matters most: each agent needs its own identity, the minimum access for its task and short-lived credentials.
  • MCP makes connecting agents to tools easy, but the protocol’s own specification says it cannot enforce security principles at protocol level. Every MCP server you connect becomes part of your supply chain.
  • Humans remain accountable. Before an agent touches real systems, someone must own it, approve its access and be able to stop it.

What makes an agent different

A chatbot produces text that a person reads and acts on. An agent decides and acts itself, often across several systems, and can sometimes create sub-agents to complete parts of a task. The NCSC notes that agents inherit known LLM risks such as prompt injection, and adds risks of their own: broader access to systems, data and tools; unpredictable behaviour when goals are interpreted in unexpected ways; problems that are harder to spot when actions happen faster than humans can review them; and actions that are harder to explain afterwards (NCSC, “Thinking carefully before adopting agentic AI”, 15 May 2026, retrieved 3 November 2026).

Put simply: with a chatbot, a successful attack changes what the model says. With an agent, it changes what the model does.

The main risks

OWASP’s agentic list is the most widely referenced map of these risks (OWASP Top 10 for Agentic Applications 2026, retrieved 3 November 2026). The table summarises the entries most relevant to a business deploying agents.

modelWhat it looks likeThe question to ask
Agent goal hijack (ASI01)Hidden instructions in an email, document or web page redirect the agent’s taskWhat untrusted content does the agent read?
Tool misuse and exploitation (ASI02)The agent uses a legitimate tool harmfully, such as deleting records or emailing data outWhich tools can it call, and with what limits?
Identity and privilege abuse (ASI03)The agent’s credentials or inherited permissions let an attacker reach far beyond the taskWhose identity does it act under, and what can that reach?
Agentic supply chain vulnerabilities (ASI04)A malicious or compromised tool, plug-in or MCP server feeds the agent bad instructions or steals dataWhere did each tool come from, and who approved it?
Unexpected code execution (ASI05)The agent generates and runs code that an attacker has influencedCan it execute code, and where?
Memory and context poisoning (ASI06)Planted content in the agent’s memory shapes its behaviour in later sessionsWhat does it remember, and who can write to that memory?
Cascading failures (ASI08)One faulty or compromised agent spreads errors through others that trust itWhich agents rely on each other’s output?
Human-agent trust exploitation (ASI09)Confident, plausible explanations persuade a person to approve a harmful actionAre approvers equipped to say no?

Agent goal hijack is prompt injection with consequences. We explain the underlying mechanism in what prompt injection is.

Permissions and tool access

Most serious agent risks come down to access. Joint guidance co-authored by the NCSC with CISA, the NSA and the national cyber security agencies of Australia, Canada and New Zealand, “Careful adoption of agentic AI services” (1 May 2026), describes a procurement agent given broad access to financial systems, email and contract repositories to reduce friction. When an attacker compromises a low-risk tool in the agent’s workflow, they inherit the agent’s excessive privileges and can modify contracts and approve payments, while the audit logs look legitimate because the trusted agent did the work (ASD’s ACSC and partners, retrieved 3 November 2026).

OWASP calls the guiding principle “least agency”: avoid giving an agent autonomy it does not need, because unnecessary agentic behaviour expands the attack surface without adding value. In practical terms, the NCSC recommends:

  • giving every agent its own unique identity, distinct from human users and other systems
  • granting only the permissions needed for the task, using credentials with the shortest possible lifetime
  • restricting which credentials an agent can see, and revoking elevated access when tasks finish
  • denying network access by default and allowing only the connections a task requires

(NCSC, “Managing the cyber risk of agentic AI”, 20 August 2026, retrieved 3 November 2026)

An agent running under a member of staff’s own account has everything that person has, which is rarely what anyone intended. Treating agents as managed non-human identities, with an owner, scoped roles and regular review, is an identity and access question as much as an AI one.

MCP: connecting agents to tools

The Model Context Protocol (MCP) is an open standard for connecting AI applications to external data sources and tools. Anthropic open-sourced it in November 2024 and donated it to the Linux Foundation’s newly formed Agentic AI Foundation in December 2025 (Linux Foundation, retrieved 3 November 2026). The current specification revision is dated 28 July 2026 (Model Context Protocol, retrieved 3 November 2026). MCP has made it quick to plug an agent into email, code repositories, databases and business applications, which is exactly why it matters for security.

Four points matter at a buyer level.

The protocol cannot enforce security for you. The MCP specification says tools represent arbitrary code execution and must be treated with caution, that tool descriptions should be considered untrusted unless they come from a trusted server, and that hosts must obtain explicit user consent before invoking any tool. It also states that MCP itself cannot enforce these principles at the protocol level; the applications that implement it must (Model Context Protocol specification, retrieved 3 November 2026).

Every MCP server is a supplier. An MCP server is software your agent trusts. A malicious or compromised one can feed the agent poisoned instructions or quietly copy what passes through it. MITRE ATLAS now documents real cases, including a poisoned MCP server used to exfiltrate email (MITRE ATLAS, retrieved 3 November 2026).

Local servers run with your user’s power. MCP’s own security guidance warns that servers installed on a user’s machine can execute commands with the client’s privileges, and that without sandboxing and consent, malicious startup commands or payloads can steal data (Model Context Protocol security best practices, retrieved 3 November 2026). Staff adding MCP servers to their own tools is a new form of shadow IT.

Credentials need care. The same guidance forbids “token passthrough”, where a server accepts tokens that were not issued to it and forwards them on, because it bypasses security controls and makes audit trails unreliable.

Useful questions for your team or supplier: which MCP servers are connected, and who approved each one? Where did each come from, and is the version pinned? What credentials does each hold? Can staff add their own?

Oversight, monitoring and accountability

The NCSC describes three oversight models: human-in-the-loop, where people approve actions before they happen; human-on-the-loop, where people monitor and can intervene; and human-out-of-the-loop, where the agent acts without review. The right choice depends on the consequences if the agent behaves unexpectedly (NCSC, retrieved 3 November 2026).

Whichever model you choose, the NCSC recommends treating agent activity as a form of user activity: logging it, including it in security monitoring and incident response, and keeping the ability to “pull the plug” immediately. It is equally clear that humans remain accountable for the decision to deploy an agent, the access it was granted, the safeguards around it and the consequences of its operation. Decide who owns each agent, who approves its access, who monitors it and who can stop it, before it connects to real systems or data.

Getting assurance before you scale

The NCSC’s advice is to start with tightly bounded pilots using clearly defined tasks, and expand only as confidence grows. In its words, “if you cannot understand, monitor or contain an agent’s actions, it is not ready for deployment” (NCSC, retrieved 3 November 2026). Before widening an agent’s scope, threat model what it could do if manipulated, then test it: AI red teaming exercises an agent’s tools and permissions the way an attacker would.

Common mistakes

  • Running agents under a staff member’s own account, so they inherit everything that person can access.
  • Granting broad access to speed up a pilot, and never narrowing it afterwards.
  • Letting staff connect MCP servers or plug-ins to work tools without review.
  • Relying on prompts that tell the agent what not to do, with no technical limits behind them.
  • Having no named owner and no tested way to stop an agent quickly.

Frequently asked questions

Should AI agents be allowed to act without human approval?

For low-risk, reversible tasks, monitored autonomy can be reasonable. For actions where the cost of error is high, such as system resets, sending data outside the network or deleting critical records, the joint agentic AI guidance recommends human review or approval checkpoints. Start with approval and relax it only when evidence supports doing so.

Are MCP servers safe to install?

Treat them like any other software that gets access to your data. Some come from established vendors; others are community projects with little scrutiny. Approve servers centrally, prefer trusted sources, pin versions and give each one only the credentials it needs.

Who is responsible when an agent gets something wrong?

Your organisation. The NCSC is explicit that a system may take an action, but humans remain accountable for deploying it, granting its access and the consequences of its operation. Agree internally who owns each agent before it goes live.

Getting started

Start with an inventory: which agents and AI assistants are in use, which tools and MCP servers they connect to, and what identity each one acts under. That alone often reveals access nobody meant to grant. Our AI and ML security team reviews and tests agentic systems, and our guide to AI security covers the wider control set. Talk to our team about the agents you are planning.

Jared Henderson avatar

Jared Henderson

Writing about cyber security and technology at Falx.

Further reading

Related insights

View all insights →
Cover for Does ISO 42001 matter for UK businesses?

10 Nov 2026

ISO 42001 is voluntary but matters if you build or sell AI products and customers ask how you govern AI. For firms only using AI tools, it can usually wait.