انتقل إلى المحتوى
العودة إلى المدونة
4 دقيقة قراءةبالإنجليزية

Five Eyes Agencies Urge Caution on Agentic AI

Security agencies from the Five Eyes alliance published joint guidance on agentic AI risks, asking organisations to put safety ahead of speed.

  • agentic AI
  • cybersecurity
  • AI governance
  • enterprise AI

On 1 May 2026, six security agencies from the Five Eyes alliance published joint guidance warning that agentic AI is being deployed too fast and with too little oversight. The agencies are CISA, the NSA and their counterparts in the UK, Canada, Australia and New Zealand, and the document is called "Careful Adoption of Agentic AI Services". It runs to 28 pages and holds more than 100 specific recommendations.

Five Eyes agencies issue guidance on secure deployment of agentic AI systems
Five Eyes agencies issue guidance on secure deployment of agentic AI systems

Agentic AI systems act on their own: they take a goal, call tools and carry out steps without a person approving each one. The guidance accepts that these systems are already at work in critical infrastructure, applying security patches, processing financial transactions and making operational decisions. Its starting assumption is blunt: "organisations should assume that agentic AI systems may behave unexpectedly."

These are the agencies responsible for protecting national security infrastructure. A joint statement from five countries about systems that are already running signals real concern about current deployments.

The risks the agencies describe

The guidance sorts the risks into five groups.

Privilege. Agents are often given far more access than their task needs. The document describes an agent authorised to install security patches whose write access is so broad that one compromise could cause widespread damage. Many organisations give agents the same access as senior administrators.

Design and configuration. Mistakes made at set-up create weaknesses before the agent does anything. Because agents run autonomously and at scale, those mistakes have a larger effect.

Behaviour. Agents may pursue their goals in ways nobody intended, and current evaluation methods cannot fully predict how they will act in new situations.

Structure. When agents depend on each other, a failure or compromise in one can cascade across the organisation.

Accountability. Agent decisions and logs are hard to audit. When something goes wrong, current tooling often cannot reconstruct the sequence of decisions that caused it.

What the guidance recommends

The agencies stop short of calling for a pause. They ask for a measured approach that puts resilience ahead of efficiency.

Deploy incrementally. Begin with clearly defined, low-risk tasks and widen an agent's responsibilities only after it has shown reliable performance.

Apply least privilege. Each agent should have a cryptographically secure identity, short-lived credentials and encrypted communications, with only the access its specific task requires.

Keep humans in control. High-impact actions should need human approval, workflows should include checkpoints, and staff should be able to interrupt or reverse an agent's actions in real time.

Defend against prompt injection. The guidance names it as a critical vulnerability: malicious inputs written to manipulate what an agent does.

Control the supply chain. Agents connect to external tools and data sources, creating what the document calls an "interconnected attack surface". That chain deserves the same scrutiny as any other software dependency.

Applying it in the UAE

The UAE is among the leading countries for AI adoption, with 54% of the population using generative AI tools, and government and enterprise are both moving quickly. At that pace, security maturity has to keep up with ambition.

Most of what the agencies ask for already exists in mature security practice, such as zero trust, defence in depth and strict access control. Those principles apply directly to agents as a new category of software. I would add a few specific habits: every agent action should be reversible before it is allowed to run autonomously, every action should be logged, and agents should be tested against prompt injection on the assumption that attackers will try it.

If your organisation already runs agents, the first step I would take is an audit of their permissions. List each agent, what it can read, what it can change, and what its task actually requires. Most organisations will find a gap, and closing it addresses the risk the agencies put first.

The full guidance is available from CISA.