Track 01

Agentic Cyber Defence

Move from alert overload to explainable, controlled action.

The challenge brief

What you are expected to build.

Agentic cyber defence uses AI agents to investigate security evidence, use tools and move an incident toward a safe, explainable response. Build one agent, a coordinated multi-agent system or an agent-enabled platform for a specific defensive task such as alert triage, vulnerability prioritisation, secure-code review or incident planning.

Minimum working outcome

A focused, testable defensive agent.

Deliver a working, testable prototype—not only a chatbot, slide deck or classifier. Your agent may use any suitable technology, including Claude, Gemini, open-source models, security APIs, automation frameworks or locally run models and tools. Demonstrate at least one realistic end-to-end use case: submit logs, code, alerts or vulnerability data; show which tools the agent chooses; reveal the evidence behind its reasoning and confidence; and produce a useful output such as a prioritised incident, proposed patch or response plan. Include human approval before consequential actions, record the agent's steps and explain limitations or possible false positives.

Your demonstration should include

01Problem & user

02Inputs & evidence

03Agent workflow

04Defensive action

05Human approval

06Result & limitations

Choose your challenge

Four project directions.

SOC triage agent

Group related alerts, reconstruct an incident timeline and recommend the next investigation step.

Vulnerability prioritiser

Combine exploit evidence, asset importance and exposure to rank what should be fixed first.

Secure-code reviewer

Find a defensible class of code flaws, validate the finding and propose a focused patch.

Incident response planner

Turn logs and a playbook into a contained response plan with explicit approval gates.

Data & references

Sample datasets—or use your own.

Use these as a starting point, bring your own permitted data or create a clearly documented synthetic dataset. Never test on systems or people without permission.

Anthropic

Your submission should prove

The agent can observe, reason and defend—responsibly.

Clear problemWorking flow or prototypeEvidence of valueHuman oversightKnown limitations