Should AI Lead Threat Hunts? A Case for Guardrails and Human Oversight

Summary
Talos author David J. Bianco argues that AI can help scale threat hunting, but should operate within narrow, carefully controlled tasks while humans retain responsibility for hunt strategy and novel analysis.
Key points
- The volume and speed of security telemetry and attacks make it increasingly difficult for human teams to keep up without automation.
- The article argues that AI can be deceived by misleading data, making trust and judgment key concerns in automated hunting.
- Bianco reframes threat hunting as a reasoning-driven process to allow for either human-led or AI-led hunts.
- He recommends starting with narrowly focused hunts that require reasoning, such as previously run hunts that teams have not fully automated.
- AI agents should have scoped, auditable, revocable credentials and only preapproved permissions and actions.
- Autonomy should increase gradually—from advisor to assistant to operator on limited tasks—to reduce the risk of harmful actions.
- Humans should decide which hunts matter and develop novel approaches; AI should execute within those bounds.
Article Details
- Defense Focus
- Use AI to conduct narrowly scoped threat hunts for security incidents missed by automated detection, while retaining human control over hunt strategy and novel analysis.
- Detection Methods
- Apply reasoning-driven analysis to security telemetry to identify incidents automated detections missed.
- Start with previously conducted hunts that have not been fully automated, where the procedure is clear but still requires judgment.
- Use analytics or machine learning to assist hunts, and allow AI to execute selected hunts within defined boundaries.
- Data Sources
- Security telemetry
- System logs
- Defensive Actions
- Give AI hunting agents scoped credentials, audit trails, revocation and rotation capabilities, and no more privilege than a junior hunter.
- Define in advance which actions an AI agent may take independently, with tighter restrictions on consequential actions such as quarantining production servers.
- Increase autonomy gradually from advisor to assistant to operator on narrowly defined tasks.
- Keep hunt priorities, strategy, and novel hunting procedures under human direction.
- Design AI hunting workflows with the risk of prompt injection and deceptive data in mind.
