Skip to content

🛡️ Cyber Security — Agentic AI

What's new

  • 2026-10-24 — Map started, from the Australian Cyber Security Centre's guidance on agentic AI.

Agentic AI — the ACSC guidance

Key takeaways from the guidance by the Australian Cyber Security Centre and partner agencies: (Cyber Security Australia)

  • Agentic AI is fundamentally different from normal GenAI because it can plan, decide, and act autonomously across tools, APIs, systems, and workflows. (Cyber Security Australia)

  • Biggest risks are:

  • Excessive privileges/access

  • Prompt injection attacks
  • Rogue agents/sub-agents
  • Cascading failures between agents
  • Hidden or non-auditable decisions
  • Third-party tool compromise
  • Emergent/unpredictable behaviour (Cyber Security Australia)

  • Traditional cybersecurity still applies:

  • Least privilege

  • Zero trust
  • Defence in depth
  • Identity management
  • Monitoring/logging
  • Human approval checkpoints (Cyber Security Australia)
  • AI security should NOT be treated separately from cybersecurity. AI systems are just another high-risk IT system integrated into enterprise infrastructure. (Cyber Security Australia)

  • The guidance strongly recommends:

  • phased rollout

  • low-risk use cases first
  • continuous testing/red teaming
  • human-in-the-loop for high-impact actions
  • strong auditability and observability (Cyber Security Australia)

What IT consultancy companies should do:

1. Stop selling “fully autonomous AI”

Position AI as:

  • supervised automation
  • co-pilot systems
  • bounded agents
  • workflow augmentation

Clients will increasingly reject “black-box autonomous AI” in regulated environments.

2. Build AI governance services immediately

Huge demand will emerge for:

  • AI risk assessments
  • AI governance frameworks
  • AI compliance mapping
  • AI architecture reviews
  • AI red teaming
  • AI audit logging
  • AI access-control design

This becomes a major consulting revenue stream.

3. Treat agentic AI like privileged infrastructure

Consultancies should architect:

  • isolated execution environments
  • least-privilege agents
  • approval gates
  • rollback mechanisms
  • sandboxing
  • tool allowlists
  • agent identity management

Basically: AI agents = semi-trusted insiders.

4. Push “human approval” workflows

Especially for:

  • finance
  • procurement
  • infrastructure changes
  • security tooling
  • customer data access
  • production deployments

High-risk actions should require explicit human signoff.

5. Create AI SOC / monitoring offerings

New opportunity:

  • agent activity monitoring
  • anomaly detection
  • prompt injection detection
  • model behaviour drift detection
  • AI incident response

“AI observability” will become a real market category.

6. Avoid over-integrated architectures

Many companies are connecting:

  • LLMs
  • APIs
  • CRMs
  • ERP systems
  • ticketing systems
  • code execution
  • email systems

…with almost no segmentation.

Consultancies should promote:

  • compartmentalisation
  • network segmentation
  • scoped permissions
  • bounded agents
  • zero-trust agent communication

7. Develop AI red-team capability

Clients will need:

  • adversarial testing
  • prompt injection testing
  • jailbreak simulations
  • tool abuse simulations
  • multi-agent failure testing

This will likely become standard for enterprise AI deployments.

8. Focus on explainability + audit trails

Future enterprise buyers will demand:

  • decision traceability
  • source attribution
  • reproducible workflows
  • action logs
  • approval records

If an AI agent changes infrastructure or approves payments, organisations need accountability.

9. Start with narrow vertical use cases

Best enterprise opportunities:

  • ticket triage
  • document classification
  • internal knowledge retrieval
  • workflow orchestration
  • SOC assistance
  • compliance automation

Avoid broad “general autonomous employee” positioning.

10. Build AI security into consulting methodology now

Companies that adapt early can position themselves as:

  • trusted AI transformation partners
  • secure AI integrators
  • AI governance specialists
  • AI operations/security providers

This guidance is basically signalling where enterprise procurement and government expectations are heading. (Cyber Security Australia)