Build vs. buy in AI security hinges on the harness

At a glance

What this is: Nullify argues that AI security build-vs-buy choices are really harness decisions, because routing, context management and guardrails shape whether agents triage findings well or mis-handle them.

Why it matters: This matters to IAM and security teams because AI agents that make or support security decisions need governed inputs, constrained outputs and clear ownership, much like other high-trust systems in identity and access programmes.

Context

The core problem is not whether AI can be used in security operations, but whether the surrounding control layer is strong enough to make its decisions reliable. In AI security, the model is only one component; routing, evidence scope, tool access and escalation rules determine whether the system behaves predictably.

That matters because AI agents increasingly sit inside security workflows that depend on judgment, evidence and bounded authority. Where the article discusses harness design, the relevant governance question is who controls the decision path, not who owns the underlying model.

Key questions

Q: What should security teams do first when AI scanners produce more findings than analysts can triage?

A: Start by separating validated, exploitable findings from unverified output and route them into distinct queues. That prevents the team from wasting scarce analyst time on low-confidence noise and lets remediation capacity focus on issues that have evidence, ownership, and business impact attached.

Q: Why do AI security systems fail when the model is strong but the harness is weak?

A: A strong model cannot compensate for missing evidence, incorrect ownership data or unsafe action rights. If the system sees the wrong inputs, it will produce a confident wrong answer. The harness determines which facts are available, which actions are allowed and when uncertainty forces escalation, so it is the real governance layer.

Q: What are the signs that AI triage is not learning properly?

A: The clearest signs are repeated analyst corrections, shifting verdicts on similar alerts, and inconsistent outcomes across shifts or model versions. If analysts keep explaining the same false positive and the system treats it as new every time, the learning layer is not retaining operational precedent. That is a governance failure, not just a model-quality issue.

Q: Should teams build or buy AI security harnesses for finding triage?

A: Buy tends to make more sense when findings span many repositories, languages and cloud accounts, or when auditors need explainable automation. Build is defensible only when the estate is small and the team can continuously maintain routing, context and guardrails. The deciding factor is ownership of the harness lifecycle, not model availability.

Technical breakdown

Why AI security harnesses need token routing

A harness routes findings to different reasoning paths based on cost, complexity and risk. Cheap first-pass models should resolve easy cases, while harder cases move to more expensive reasoning only when the evidence justifies it. Without routing, a team spends premium inference on noise and still fails to improve decision quality. The article’s example of a triage stage using 97.5% fewer input tokens shows that architecture, not raw model strength, drives efficiency. Routing also has to handle timeouts, provider failures and shifting model economics without breaking the workflow. See the AI Security Platform Buyer's Guide.

Practical implication: Define decision tiers for findings before adding more model capacity.

How context management changes AI security accuracy

Context management controls what evidence the agent can see before deciding. A snippet alone is often insufficient, because the deciding fact may live in a caller, service map, ownership record or policy rule. When the agent sees the wrong slice of evidence, even a strong model can produce the wrong answer. The design challenge is not just retrieval, but freshness: the evidence must stay aligned with code, cloud resources and organisational ownership as they change. In practice, context management is a governance control as much as a technical one. See the AI Security Maturity Model.

Practical implication: Bind triage decisions to the evidence set that actually determines risk.

Why guardrails matter more when agents can act

An agent that can dismiss findings or open pull requests needs explicit boundaries for action and escalation. The failure modes are asymmetric: a false dismissal hides a real issue, while a false escalation only creates review overhead. That means dismissal should require affirmative evidence, uncertainty should route to deeper review, and generated fixes should be validated before humans spend time on them. This is the point where AI security becomes operational governance. The issue is not whether the system can act, but whether it can be prevented from acting on weak evidence. See the AI Agent Identity Security Buyer's Guide.

Practical implication: Set a much higher evidence bar for dismissal than for escalation.

Threat narrative

Attacker objective: The practical objective is to hide or downgrade real security findings so they do not receive human review or remediation.

1. Entry begins when an AI agent or harness ingests security findings and starts making triage decisions based on incomplete or poorly routed evidence.
2. Escalation occurs when the system is allowed to dismiss, reroute or generate fixes without a strong evidence threshold or validated context.
3. Impact follows when a real vulnerability is wrongly dismissed, leaving a security issue hidden inside an automated workflow rather than escalated for review.

Breaches seen in the wild

• 12,000 secrets in LLM training data: Truffle Security found 11,908 live API keys and passwords hard-coded in web pages captured by Common Crawl, a dataset used to train LLMs.
• Microsoft SAS token exposure 2023: An over-permissive Azure SAS token in a Microsoft AI GitHub repo exposed 38TB, including workstation backups and Teams messages, for 3 years.

Read and download The State of NHI & AI Agent Breach Report 2026, covering 200+ breaches impacting Non-Human Identities including AI Agents.

NHI Mgmt Group analysis

AI security is now a harness governance problem, not a model-selection problem. The article’s central argument is correct: once the same frontier or open-weight models are available to everyone, the differentiator becomes how decisions are routed, constrained and audited. That shifts value from model novelty to operational design, which is exactly where governance failures usually emerge. Practitioners should treat the harness as the control plane for AI decisions.

Context quality is the hidden determinant of AI security accuracy. A model that sees the wrong evidence will make a confident wrong call, even if its reasoning is strong. That is a familiar security pattern in new form: decision quality depends on identity, ownership and environmental facts being attached to the finding before the model acts. Where AI agents support security triage, third-party AI risk principles also matter because the decision source chain becomes part of the control surface.

AI triage needs asymmetric controls because dismissal is the dangerous action. A false escalation costs time, but a false dismissal creates unreviewed exposure. That means evidence thresholds, uncertainty handling and validation gates should be stricter for dismissals than for escalations. In identity terms, this is the same logic as limiting authority until trust conditions are proven.

Build decisions create long-term governance debt unless ownership is explicit. If a team builds its own harness, it is also taking on model re-evaluation, routing changes, evidence policy and auditability as permanent responsibilities. That makes internal ownership a programme decision, not a tooling preference. Practitioners should assume the harness will outlive the first model it runs.

From our research library:

• Only 23% of IT leaders were very confident in their organisation's ability to manage security and governance for GenAI deployments, according to a 2025 Gartner survey of 360 IT leaders.
• Read next: AI Security Platform Buyer's Guide

What this signals

AI security operations are moving toward policy-driven decision systems. The practical question is no longer whether a model can reason over findings, but whether the organisation can govern what it sees, what it may change and when it must defer. That pushes AI security closer to access control and workflow governance than to simple model deployment.

Decision rights will become the differentiator in AI-assisted security workflows. Teams that can separate cheap screening, expensive reasoning and human escalation will get better economics and lower risk. The organisations that fail will usually have one of two problems: too much automation authority or too little evidence discipline.

Context, not scale, will define the next AI security maturity gap. As more teams adopt AI triage, the gap will widen between those that can bind findings to ownership, assets and policy, and those that only feed snippets to a model. That is a governance problem disguised as a tooling choice.

For practitioners

• Define routing rules before adding more models Map finding types to the cheapest reliable model stage, then reserve expensive reasoning for cases that truly need it. Document timeout and fallback behaviour so the pipeline still works when a provider fails.
• Attach decision-grade context to every finding Require the agent to see the repository state, owning service, related cloud resource and applicable policy before it can decide. If the evidence is not available, the finding should advance, not be guessed away.
• Gate dismissals on affirmative evidence Treat dismissal as a high-risk action and require proof that the issue is not exploitable before closing it. Uncertainty should route to human review, not to an automated no.
• Measure false-dismissal performance separately Track how often the harness suppresses real findings, not just how many items it processes. That metric is the clearest signal of whether automation is reducing risk or simply shifting it.

Key takeaways

• AI security build-vs-buy decisions are really decisions about governance of the decision layer, not about which model to call.
• Routing, context and guardrails determine whether automation reduces triage cost or simply accelerates bad judgments.
• The most important control is the evidence standard for dismissal, because that is where hidden security failures can slip through.

Related NHI Mgmt Group resources: AI Security Platform Buyer's Guide · AI Security Maturity Model · AI Agent Identity Security Buyer's Guide · Third-Party AI Risk Guide

Key terms

• AI Security Harness: The harness is the surrounding control layer that routes tasks, manages context, sets tool access and governs outputs for an AI system. In security workflows, it often matters more than the model itself because it determines what evidence is visible, what actions are allowed and when humans are involved.
• Decision-grade context: Security information that is rich enough to change an operational decision, not just inform an analyst. In practice, this means telemetry and enrichment that can alter access, response, escalation, or containment based on identity, asset, and threat relevance.
• False Dismissal: A false dismissal happens when a security system wrongly closes or suppresses a real issue. It is more dangerous than a false escalation because the problem may remain hidden from human review, allowing vulnerability or exposure to persist inside an automated workflow.
• Routing Policy: Routing policy is the set of rules that determines which model, tier, or fallback path handles each AI request. When the policy is weak, organisations can under-protect high-stakes work or over-spend on routine tasks.

Deepen your knowledge

The NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance and machine identity security for practitioners who need to connect access decisions to operational control. It helps identity and security teams evaluate where human oversight, lifecycle control and delegated authority belong in modern programmes.