Cybersecurity & Data Protection · IACT Training Blog
Through 2026, Irish businesses have rushed to deploy AI agents, autonomous systems with access to email, files, APIs and line-of-business apps. That access is exactly what attackers want. Prompt injection, where malicious instructions are smuggled into the data an agent reads, has become one of the highest-risk threats of the year precisely because it is scalable and hard to detect.
Why agents change the threat model
A chatbot that only answers questions has a small blast radius. An agent that can read a mailbox, query a database and send messages has a large one. When an agent is implicitly trusted with privileged access, a single manipulated instruction can move money, leak data or change configurations. Industry analysts reported AI-related attacks rising sharply through early 2026, and over 90% of breaches still begin with some form of social engineering, now aimed at machines as well as people.
Prompt injection has evolved from crude jailbreak attempts into multi-step campaigns. In ‘salami slicing’ attacks each individual prompt looks harmless, but the cumulative effect gradually shifts the agent’s understanding of its goals until it does something it never should.
The main attack patterns
- Direct prompt injection: a user types instructions that override the agent’s system rules.
- Indirect prompt injection: malicious text hidden in a web page, document or email that the agent later reads and obeys.
- Tool and data poisoning: tampering with the sources or connectors an agent trusts.
- Excessive agency: an agent granted broader permissions than its task requires, amplifying any single compromise.
Practical defences
Treat every agent as you would a new employee with system access. Apply least privilege so it can only touch what its job needs. Keep a human in the loop for high-impact actions such as payments, deletions or external sends. Separate trusted instructions from untrusted content, and never let data the agent retrieves silently become commands.
- Scope credentials tightly and rotate them; avoid shared admin tokens.
- Log and monitor agent actions so anomalous behaviour is visible.
- Validate and sanitise inputs, and constrain outputs that trigger real-world actions.
- Test agents adversarially before go-live, the way you would penetration-test an app.
Awareness is still the front line
Technology controls matter, but people remain decisive. Effective 2026 security awareness programmes have moved beyond a once-a-year e-learning module to continuous, behaviour-based training with realistic simulations. Staff who understand how AI agents can be manipulated are far less likely to paste a suspicious document into one, or to approve an action they do not understand. For Irish organisations, this also supports GDPR accountability obligations around data handling.
Train with IACT
IACT delivers cybersecurity awareness and data-protection training for Irish teams, including AI risk, phishing simulation and GDPR. Equip your people to be the strongest control.