← Back to blog

AI Agent Security Risks: What IT Teams Must Control

November 17, 2026
AI Agent Security Risks: What IT Teams Must Control

The biggest AI agent security risks are prompt injection, agents holding more access than they need, and agents nobody can account for. Ireland's National Cyber Security Centre names prompt injection, excessive agency and compromised agents as the core risks. The consequences it lists are unauthorised actions, data exfiltration, lateral movement and operational disruption. All three get worse when you cannot see which agents exist, who built them and what they can reach. In Microsoft 365, that visibility is the first control to put in place, and it is the one our platform, Argus, audits every day.

Secureai
See Which AI Agents Need Attention
Argus audits Microsoft 365 AI agents daily, links each to its creator and owner, and verifies fixes to help teams manage security risks.
Visit

TL;DR

  • Three risks dominate: prompt injection, excessive agency (too many permissions, connections or autonomy) and compromised agents that act on an attacker's behalf.
  • Most of the damage comes from what an agent is allowed to do, not from the model itself. Least privilege is the strongest single broad control in a 2026 peer-reviewed study.
  • You cannot govern what you cannot list. Only 37% of organisations keep a complete, current inventory of agents and their authorised actions.
  • Every agent needs a named owner, a purpose and a fix that is verified, not just ticketed.
  • Argus audits Microsoft 365 AI agents daily, links each to its creator and owner, ranks the risks and verifies fixes.

Table of Contents

What an AI agent changes about your risk

A chatbot answers questions. An agent acts: it reads files, calls connected systems, starts workflows and makes decisions at machine speed. PwC Ireland's Len McAuliffe, quoted in The Irish Times, describes the risk as coming from agents' data access, system interaction, workflow initiation and rapid decision-making. Broad permissions, weak oversight or manipulated instructions can escalate the consequences quickly.

The UK NCSC makes a useful point in its May 2026 agentic AI guidance. Agentic systems inherit known risks such as jailbreaking and prompt injection, and the added autonomy widens the attack surface. Access control, monitoring, incident response and accountability still matter. What changes is that behavior is harder to predict, test and govern.

For a security team in a Microsoft 365 tenant, this means the risk is rarely one dramatic exploit. It is a slow accumulation of agents that people across the business created, shared and connected, each with its own permissions.

The main AI agent security risks

Prompt injection

An AI model may not reliably tell a legitimate instruction from malicious content. The NCSC's assessment explains that an attacker can plant instructions in a website, file or email that the agent reads. The result can be goal hijacking, data extraction, leaked prompts or infrastructure details, or commands run in connected systems. See the NCSC AI Cyber Security Risk Assessment.

You cannot fully remove this risk by editing a prompt. What you can do is limit what a hijacked agent is able to reach. An injected instruction is only as damaging as the permissions and connections behind the agent.

Excessive agency and over-permissioning

Excessive agency means an agent has more functions, permissions or autonomy than its task requires. The NCSC links it to unauthorised access, data breaches and operational disruption. A typical example is an agent built to summarize a team's documents that is also shared with the whole organization, or connected to a system holding sensitive records.

This is the risk security teams can act on most directly, because it is a configuration question: who can use the agent, what it connects to and what it can do there.

Data exposure through connections

Every connection is a path for data to leave. Agents connected to sensitive sources, or shared without restriction, can expose information to people who should never see it. Prompt injection is one route to exfiltration. A careless sharing setting is another, and it needs no attacker at all.

Compromised agents and lateral movement

The NCSC states that a compromised agent can execute unauthorised actions, move laterally, exfiltrate data and cause widespread disruption. An agent that holds standing credentials and broad reach gives an attacker a ready-made foothold that already looks like legitimate activity.

Autonomy without a checkpoint

The Cloud Security Alliance's 2026 whitepaper, Autonomous by Design, Uncontrolled in Practice, advises you to inventory every agent with standing authority to delete, modify or exfiltrate data. It also says to identify which of them can act without per-action human confirmation. It treats settings that remove approval checkpoints as governed changes, not casual toggles.

Agents nobody owns

The quietest risk is orphaned agents: built by someone who has since changed roles, still running, still connected, with no one answerable for them. The same Cloud Security Alliance paper recommends giving every agent a documented owner, a purpose and a decommissioning process. For compliance officers this is the heart of the matter. If an auditor asks who approved an agent's access to a data source, "we are not sure who built it" is not an answer.

Why visibility comes first

Most controls assume you already know what exists. In practice, many organizations do not. The 2026 Power without control survey found that only 37% of organisations maintain a complete, up-to-date inventory of agents and their authorised actions. The same share lacks full visibility into all permissions, decision rights or actions. Only a quarter regularly report mean time to recover to the board or C-suite.

Survey figures on AI agent visibility gaps

The survey also found public-sector organisations especially exposed: 18% reported robust rollback capabilities and 11% completed regular recovery-time reporting. If you work in an Irish public body, this is the gap the NCSC's public-sector guidance is aimed at.

The survey quotes one participant who argues that only with observability can an organisation understand how an agent operates, what data it ingests and what decisions it makes. That is the precondition for reversing a bad outcome.

A workable visibility baseline answers five questions for every agent:

  1. Who created it, and who owns it now?
  2. What is it connected to?
  3. Who can use or share it?
  4. What can it do in those systems?
  5. What changed since yesterday?

Controls that reduce the risk

The research agrees on a short list. A peer-reviewed April 2026 study of tool-using agents tested several defences and found these results:

ControlWhat the study found
Least-privilege tool designThe strongest single broad control
Human-in-the-loop approvalsSharply reduce high-impact actions and exports, but weaken when users get used to approving
Retrieval allowlistingNearly eliminates indirect prompt injection, with little effect on other channels
Layered stackPerformed best overall

Its conclusion is that agent safety is an architectural and operational problem. Which defense fits depends on the threat. No single switch covers everything.

The UK NCSC's practical steps point the same way: least privilege, limited scope, no long-lived credentials, secure defaults, supply-chain management, behavior monitoring, threat modelling and incident planning. The Cloud Security Alliance adds enforcement outside the model, through policy engines or identity gateways issuing narrowly scoped, short-lived credentials. It also asks for logging of every tool call, with principal, arguments and outcome, to a tamper-proof sink.

Notice what most of these have in common. They depend on knowing which agents exist, how they are configured and whether the configuration has drifted. Controls decay when nobody checks them again.

What the rules in Ireland expect

This section is for compliance officers who need to map agent risk to obligations. The EU AI Act applies directly in Ireland, and, according to a legal overview of Irish AI law, its phased application began on 2 February 2025. The same overview reports that the General Scheme of the AI Bill, published on 4 February 2026, sets out an Irish enforcement framework with fines of up to 7% of global turnover for the most serious breaches. Because timelines and proposed amendments can move, check the current position on the official Department of Enterprise page below.

The NCSC states that the EU AI Act, NIS2, GDPR and national frameworks together set obligations for public-sector bodies deploying AI. It also says high-risk AI systems must be designed to reach an appropriate level of cybersecurity across their lifecycle (see its risk assessment). It recommends the Cyber Fundamentals (CyFun) framework, grounded in NIST CSF v2.0, to help entities meet NIS2, with certification expected in Ireland in 2027.

The Department of Enterprise's guidance, The EU AI Act and my organisation, gives a practical order of work:

  • Map where AI is used.
  • Determine the risk category and check for prohibited practices.
  • Put governance and oversight in place.
  • Ensure transparency and documentation.
  • Build AI literacy and work with trusted suppliers.
  • Monitor updates from national and EU authorities.

The NCSC's own baseline actions begin the same way: establish governance, with roles, accountabilities and usage policies, before deployment, then identify and prioritise at-risk assets. An academic analysis in Springer notes that the AI Act and the Cyber Resilience Act state outcome-oriented obligations rather than implementable safeguards. That leaves your team to translate them into concrete controls. An up-to-date agent inventory with named owners is the evidence base nearly every one of these frameworks assumes.

A practical sequence for a Microsoft 365 tenant

If you are starting from nothing, this order produces the most risk reduction per week of effort.

  1. Build the inventory. List every agent in the tenant: name, creator, current owner, connections, sharing scope. Re-run it often, because agents appear whenever someone builds one.
  2. Assign accountability. Give every agent a named person and a purpose. Agents without an owner get reviewed first, then retired if no one claims them.
  3. Rank by exposure. Start with unrestricted sharing and connections to sensitive data. These are the settings an attacker or an honest mistake can exploit without any sophistication.
  4. Cut permissions to the task. Remove access the agent's purpose does not need. Prefer short-lived, narrowly scoped credentials, as both the UK NCSC and the Cloud Security Alliance advise.
  5. Gate destructive actions. Require human confirmation for deleting, modifying or exporting data, and treat any change that removes that checkpoint as a governed change.
  6. Verify, then monitor. A closed ticket is not proof that the setting changed. Confirm the fix, then watch for drift.
  7. Fold it into existing processes. The NCSC asks organizations to integrate AI risk into existing cyber governance and incident response, not to build a parallel system.

Steps 1 to 3 and step 6 are where manual effort breaks down. A spreadsheet is stale the day after it is written, and agents are created by people across departments, not only by IT.

How Argus covers the visibility gap

Argus is our security platform for Microsoft 365 environments. It audits every AI agent in the tenant daily and records who created each agent, where it connects and who holds permissions. It ties each agent to its creator and current owner, so accountability is on record, which matters in regulated industries.

How Argus covers the visibility gap

It identifies and ranks risks such as unrestricted sharing and sensitive connections, and it alerts you in real time when something risky changes. Where a fix is made, Argus verifies it and gives you evidence that the issue is resolved, rather than leaving a ticket marked done. Findings can feed the tools your team already uses, including Jira and ServiceNow. Argus does not make changes to your tenant and does not require additional Microsoft licenses.

For a CISO, that is a daily, evidence-backed answer to "what agents do we have and what can they reach?" For a compliance officer, it is a record of who owns each one. You can see what Argus checks in Copilot Studio agents, read how it works on the Argus page, or look at customer stories.

Secureai

To see Argus against your own tenant, start at secureai.ie.

How to explain agent risk to business managers

Business managers do not need the threat taxonomy. They need three messages:

  • Agents act, not just answer. An agent can open files and trigger workflows with the permissions it was given.
  • Convenience creates exposure. Sharing an agent widely or connecting it to a sensitive source is quick to do and is the most common route to data exposure.
  • Someone must be answerable. Each agent should have a named owner in the business, because the security team cannot judge on its own whether a given connection is justified.

This also changes how you report. Instead of counting agents, report the number of agents without an owner, the number with unrestricted sharing, and how long fixes take to be verified. These figures can move on a weekly basis and show progress a board can follow. The survey above found that only a quarter of organisations regularly report recovery time to the board or C-suite, so even basic reporting sets you apart.

Common mistakes

  • Treating it as a model problem. The risk lives in permissions, connections and sharing, which you control.
  • Relying on approvals alone. The 2026 study found human approval degrades as users get used to it, so pair it with least privilege.
  • Auditing once. The NCSC stresses that AI risk must be reassessed as models update and threats develop, so the audit has to repeat.
  • Skipping verification. A fix that is assumed, not confirmed, is a risk you believe is closed.
  • Leaving owners blank. An orphaned agent has nobody to question when its behavior changes.

FAQ

What are the top 3 AI security risks?

The NCSC identifies prompt injection, excessive agency and compromised agents. Between them they lead to unauthorised actions, data exfiltration, lateral movement and operational disruption. In practice, the second one, excess permissions and connections, is the easiest to find and fix.

Is it safe to have an AI agent in my organization?

It can be, provided security continues through the whole lifecycle. The NCSC supports AI adoption on that basis, and it asks for governance, risk assessment of at-risk assets and regular reassessment. An agent with a named owner, limited permissions and daily monitoring is a far smaller risk than an unmanaged one.

What is the difference between prompt injection and excessive agency?

Prompt injection is the attack: malicious content that manipulates what the agent does. Excessive agency is the weakness that makes the attack costly: the agent holds more access or autonomy than it needs. Fixing the second limits the damage from the first.

How often should AI agents be audited?

Agents are created and changed continuously, so a one-off review ages quickly. The NCSC says AI risk must be regularly reassessed, and Argus audits Microsoft 365 agents daily. Choose a frequency that matches how fast your teams build agents.

Who should own an AI agent?

A named person in the team that depends on it, recorded alongside its purpose. The Cloud Security Alliance recommends a documented owner, purpose and decommissioning process for every agent. The security team supports and audits, but the business owner justifies the access.

Sources