A Guide to Adopting Tiered AI Defense

The open letter on collective cyber defense lays out a decision architecture that many security teams haven’t yet formalized:

“Use capable, lower-cost models for broad coverage, and apply frontier capabilities to the hardest problems.”

This isn’t just a cost recommendation. It’s a security architecture in its own right: broad coverage with fast, cheap models, depth with frontier models only where it’s actually needed. This guide gives you a concrete framework for deciding which task belongs at which tier.

Why “use the biggest model for everything” is a design flaw

Running the most powerful available model on every alert your SIEM generates is slow, expensive, and — counterintuitively — less secure: it eats your budget and your latency exactly when you need speed the most. The right architecture is a tiered triage pipeline, similar to how a human SOC escalates from an L1 to an L3 analyst.

Tier 1 — Low-cost models for mass triage

Tasks that belong here: high volume, known patterns, low risk from a costly false negative.

• Initial alert classification (is this noise, a likely false positive, or does it need review?).
• Log normalization and enrichment.
• Detection of known phishing patterns (lookalike domains, suspicious headers).
• Automated ticket summaries for the human analyst.

Characteristics to look for in the model for this tier: low latency, low cost per token, good classification performance (no need for deep reasoning).

python

# Simplified example: alert triage with a low-cost model
def triage_alert(alert):
    response = client.messages.create(
        model="low-cost-model",
        max_tokens=200,
        messages=[{
            "role": "user",
            "content": f"Classify this alert as NOISE, REVIEW, or CRITICAL. Alert: {alert}"
        }]
    )
    return parse_classification(response)

Tier 2 — Conditional escalation

Between the low-cost tier and the frontier tier you need an explicit decision rule, not “when in doubt.” Example escalation triggers:

• Tier 1 classification came back “REVIEW” or “CRITICAL.”
• The asset involved is on your essential infrastructure list (defined in the technical debt audit post).
• The pattern doesn’t match any known signature (possible novel attack).
• There are signs the attack itself is using generative AI (more on this below).

Tier 3 — Frontier models for the hard cases

Reserve the most capable model for what actually needs it:

• Analysis of novel or polymorphic malware, which requires reasoning about behavior, not just signatures.
• Investigation of complex incidents with multiple sources of evidence to correlate.
• Red teaming and simulation of AI-assisted attacks (see the dedicated post on this topic).
• Generation of response playbooks for scenarios with no precedent in your organization.
• Review of AI-generated code in high-criticality components (see the dedicated post).

Designing the full pipeline

Incoming alert
      │
      ▼
[Tier 1: Low-cost model] ── NOISE ──► Archive (with random sampling for QA)
      │
   REVIEW / CRITICAL / unknown pattern
      │
      ▼
[Tier 2: Escalation rules] ── doesn't meet criteria ──► Human L1 analyst
      │
   meets escalation criteria
      │
      ▼
[Tier 3: Frontier model] ──► Deep analysis + recommendation
      │
      ▼
Human L2/L3 analyst (final decision — the model doesn't autonomously execute actions in critical cases)

Important point: at tier 3, the model recommends, it doesn’t autonomously execute irreversible actions (isolating a production host, mass credential revocation). That’s the human’s role, or automation with explicit approval — something directly tied to the agentic identity traceability we cover in another post in this series.

Measuring whether the architecture is working

• Cost per alert processed: should drop significantly versus using frontier for everything.
• Time to containment on critical cases: shouldn’t get worse — tier 3 should kick in fast when it matters.
• False negative rate at tier 1: audit periodically with manual sampling to confirm the low-cost model isn’t letting things through that should have escalated.

Implementation checklist

• ☐ Map your current security tasks onto the 3 tiers
• ☐ Define explicit escalation rules (not discretionary)
• ☐ Implement tier 1 with a low-cost model for high-volume triage
• ☐ Reserve the frontier model only for cases meeting escalation criteria
• ☐ Ensure irreversible actions require human approval or audited automation
• ☐ Measure cost, time to containment, and false negative rate periodically

Part of a series on how to put into practice the principles from OpenAI’s open letter on collective cyber defense (August 2026). Back to the full guide.