Field guideUnclassified · for security leadership
- Guide
- BM-RG-01
- Prepared for
- CISOs and SOC leads
- Updated
- Reading time
- 14 min
Adaptive AI for breach response.
Abstract
Containment is now decided in minutes, and a human-only triage queue cannot keep that pace. Autonomous agents can read every alert, but an agent that acts without a signature is a new way to fail.
This guide sets out where agents help, how attackers go after them, and a five-stage path to deploying them with a person’s approval on every consequential action.
- Every alert enters; the agent narrows them
- The agent investigates and drafts a plan
- Nothing runs until a person signs
1.The queue is the vulnerability
A modern security operations centre watches infrastructure, SaaS event streams, identity signals and threat intelligence at once. Hybrid and multi-cloud estates have multiplied each of those feeds, and even strong teams face queues of thousands of alerts a day. The cost shows up as dwell time: the interval between the first compromise and the moment someone notices.
Traditional response assumes a person reads each alert, decides whether it matters and starts the investigation. That assumes constant availability, perfect prioritisation and unbroken attention. In practice analysts are tired, interrupted and triaging by severity labels that attackers have learned to stay under. Low-priority queues are where lateral movement hides, and nights and weekends are when it happens.
Regulators, boards and insurers now ask for the evidence: how long did detection take, how long did containment take, and who decided. A queue cannot answer that well.
2.What an agent changes, and what attackers do about it
Assistant-style automation waits for an analyst to ask. An agent starts the work itself: it enriches the alert, correlates it with other telemetry and threat intelligence, follows a hypothesis until it confirms malicious activity or clears the alert, and drafts the response. It does this for every alert, not only the ones a person reached. The noise that used to bury an attacker’s signal no longer works.
Attackers adapt. They aim prompt injection at the agent’s inputs, plant misleading context in tickets and logs, and generate benign-looking traffic to steer its conclusions. An agent without a readable decision log, a stated reason for each step and hard escalation limits will either act too aggressively or hesitate when it should not.
Human-in-the-loop design resolves that tension. The agent is the first responder; a person keeps authority over anything with consequences. Whether autonomy is an advantage or a liability comes down to that governance.
3.Five stages from readiness to routine
Agents belong in an incident response programme the way any new authority does: assessed, scoped, supervised and measured. We use five stages.
- Readiness assessment. Count your alerts and measure today’s time to investigate and contain, so you can prove a change later. Map the telemetry sources (endpoint, network, identity, SaaS, cloud) and write down how analysts actually work a case. List the regulatory and data-residency rules that limit what an agent may touch.
- Choose the autonomy tier per workflow. Decide, workflow by workflow, whether the agent assists (a person leads), augments (shared work) or acts (the agent leads within limits). It must read from and write to your SIEM, SOAR, case management and ticketing, or it becomes one more console.
- Design the human in the loop. Set the thresholds that force escalation, the chain of approvers and the rollback for every action. Each agent step writes to an append-only log, and each proposal arrives with its evidence, its confidence and a plain-language reason a person can check.
- Fit it into the runbooks. Agree case tags, alert states and naming across tools. Feed every human decision back to the agent so its proposals improve. Train the people who will sign, not just the people who will configure.
- Measure and adjust. Track detection latency, time to containment, automation coverage, analyst hours returned and false-positive and false-negative rates. After each incident, decide whether a workflow earns more autonomy or less.
The decision logic itself should be short enough to read aloud in a review. This sketch is the shape we start from, before asset criticality, data classification and business-impact thresholds are added:
if alert.severity > threshold and agent.confidence > 0.9:
finding = agent.investigate(alert)
if finding.confirms_compromise:
plan = agent.draft_containment(finding)
approval_queue.submit(plan, evidence=finding.evidence)
# Nothing runs until a named person signs the plan.4.Guardrails that keep a person on the record
- A person confirms any containment that touches regulated workloads, executive accounts or revenue-carrying production paths.
- Destructive actions (revoking credentials, segmenting networks, shutting systems down) need two approvers.
- Every agent proposal carries its reason, the evidence behind it and the steps to roll it back.
- Audit on a schedule for model drift, actions that misfired and attempts to manipulate the agent.
None of this is purely a technology project. Legal, business continuity and communications sit on the approval chain for the actions that affect them, so the agent’s limits match the organisation’s appetite for risk.
5.Worked scenario: a federation misconfiguration
An illustrative case, built from patterns we see rather than a single client. A multinational runs on-premises data centres, three public clouds and thousands of SaaS integrations. Its SOC sees about ten thousand alerts a day; analysts clear the critical queue in roughly ninety-five minutes and low-severity alerts sit for whole shifts. An attacker abuses a misconfigured identity federation, establishes persistence and moves laterally through alerts nobody opened.
With an agent triaging every alert, the lateral movement is flagged within minutes as anomalous service account use. The agent ties it to the recent federation change and drafts a plan: isolate the affected hosts, reset the compromised credentials, force re-authentication on critical systems. The plan reaches the on-call incident commander with its evidence, its confidence and its predicted business impact. The commander signs; containment completes in about eleven minutes.
The speed matters less than what it leaves behind: analysts freed for hunting, and an audit trail in which every agent step, every piece of evidence and every human decision is recorded and timestamped.
6.What leadership owns
- Investment discipline. Fund a pilot with stated gates: time to contain, analyst capacity returned, risk reduced. Release the next budget when the gate is met.
- Metrics. Track time to investigate, contain and remediate, the share of alerts fully triaged by agents, how often people override them and the false-positive rate.
- Governance. Name an oversight group, the escalation thresholds and the documentation standard. Expect regulators to ask how automated decisions were made and who approved them.
- Adversarial resilience. Model how inputs could be poisoned or prompts hijacked, then validate inputs, sandbox execution and red-team the agent continuously.
- People. Train analysts to supervise agents and read confidence. Put the hours returned into threat hunting, deception and resilience testing.
Phase 1
Pilot human-approved triage
One agent, a narrow slice of alerts. Measure time to investigate and the false-positive rate, and collect analyst feedback to tune the confidence thresholds.
Phase 2
Triage everywhere, containment proposed
Extend coverage to every telemetry source. The agent drafts containment plans; a person approves anything with high impact before it runs.
Phase 3
Learning from every signature
Human decisions train the agent, deception sensors feed it, and playbooks span domains so the agent can anticipate an attacker’s next pivot. The approval gate stays.
7.Verdict
Autonomous agents are ready to be first responders. They shorten detection and do not tire, but they do not carry context, ethics or business judgement. Those stay with people. Organisations that put agents to work behind an approval gate now will contain faster and be able to show exactly how they did it.