How should organisations adapt their security strategy as AI becomes part of both attack and defence?

Organisations should treat AI as an operational capability on both sides of the security equation. Defenders can use AI to improve detection, triage, and response, while attackers can use it to scale reconnaissance and social engineering. The practical response is stronger identity controls, faster verification, and diverse defensive perspectives so teams can spot and disrupt AI-assisted abuse sooner.

How AI Changes the Threat Model for Everyday Security Work

As AI becomes embedded in both offensive and defensive workflows, the security strategy needs to shift from static perimeter thinking to faster, evidence-based decision making. Attackers can use AI to accelerate reconnaissance, content generation, and social engineering at scale, while defenders can use it to compress detection and triage cycles. The practical implication is that organisations must assume greater speed, more volume, and more convincing deception on both sides.

A useful way to frame the change is that AI lowers the cost of iteration. That means the security function no longer wins by simply adding more manual review, it wins by improving signal quality, reducing time-to-verify, and making it harder for an attacker to move from contact to compromise without being challenged.

That is why identity, verification, and trust boundaries become more important, not less. AI-assisted abuse often succeeds when the defender accepts a message, token, workflow, or request too quickly, and the best countermeasure is usually to make the environment harder to impersonate and easier to validate. NHIMG’s Ultimate Guide to Non-Human Identities is a strong reference point here because strong identity hygiene remains one of the few controls that still scales when attackers automate.

What Defenders Should Adapt First

The first adaptation is to treat AI as an operational capability, not just a tool. If defenders use AI only for productivity gains, they miss the chance to use it for faster enrichment, prioritisation, and pattern detection. The second adaptation is to diversify defensive perspectives, because AI-generated phishing, deepfake voice, synthetic support interactions, and automated recon can defeat single-channel verification.

Move high-risk approvals to channels that are harder to spoof, especially where money movement, privilege changes, or access resets are involved.
Use AI to triage large queues, but keep escalation rules tied to observable evidence rather than model confidence alone.
Review which actions can be completed by automation without human challenge, and narrow that set where the blast radius is high.

Controls that manage secrets, tokens, and access paths matter more in this model because AI-assisted attackers are often looking for the fastest path from a single foothold to wider reach. The same applies to stale credentials and over-privileged accounts, which create easy leverage points for rapid abuse. The 52 NHI Breaches Analysis is relevant because it shows how compromised machine-facing access can become a force multiplier once abuse begins.

For broader operational tuning, defenders should also align playbooks to known defensive patterns and attacker tradecraft. MITRE D3FEND helps teams think in terms of countermeasures, while CISA advisories help teams keep pace with active threat behaviour and current abuse patterns.

Risk and Threat Considerations

The main risk is not that AI creates entirely new attack classes, but that it makes familiar ones faster, cheaper, and more persuasive. Social engineering becomes more scalable, reconnaissance becomes more automated, and defenders face more false confidence if they over-trust AI output without independent validation.

Failure mechanism: Attackers use AI to accelerate lure generation, impersonation, and target selection, then exploit weak verification, excessive access, or slow incident handling to turn one successful interaction into broader compromise.

Impact: Organisations can see higher-volume phishing, faster credential abuse, more convincing pretexting, and shorter time windows to detect and stop an intrusion. Where secrets, tokens, or service accounts are exposed, the resulting access can be exploited at machine speed.

NHIMG’s research on DeepSeek breach is a useful illustration of how log exposure and sensitive keys can turn an AI-adjacent incident into a broader identity and data problem. For attackers, exposed secrets remain one of the highest-value shortcuts because they bypass much of the normal user verification chain.

Anthropic’s report on the first AI-orchestrated cyber espionage campaign and the MITRE ATLAS adversarial AI threat matrix are both useful for understanding how AI changes attacker sequencing, especially where tool misuse, autonomous recon, and rapid iteration are involved.

Standards & Framework Alignment

This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.

OWASP Non-Human Identity Top 10 and MITRE ATT&CK address the attack and risk surface, while CIS Controls v8 and NIST Zero Trust (SP 800-207) set the governance and control requirements practitioners need to meet.

Framework	Control / Reference	Relevance
OWASP Non-Human Identity Top 10	NHI-01 — Secrets and Credential Management	AI-assisted abuse often pivots through exposed secrets and access material.
	NHI-02 — Identity and Access Management	Stronger identity controls are central when AI raises impersonation and abuse speed.
	NHI-06 — Visibility and Monitoring	Defenders need faster detection of AI-driven abuse and unusual access paths.
Recommendation — Audit and rotate exposed secrets to reduce AI-amplified credential abuse. Enforce least privilege and tighter verification for high-impact access. Improve telemetry on identity actions so AI-assisted abuse is easier to spot.
MITRE ATT&CK	TA0006 — Credential Access	AI can scale reconnaissance and credential harvesting before deeper intrusion.
	TA0001 — Initial Access	AI strengthens phishing and impersonation used to gain first foothold.
	T1589 — Gather Victim Identity Information	AI accelerates target profiling for more convincing pretexts and lures.
Recommendation — Hunt for credential access activity and block harvesting workflows early. Prioritise controls that reduce initial-access success from social engineering. Detect automated reconnaissance that collects identity data for impersonation.
CIS Controls v8	5 — Account Management	Stronger account governance limits abuse of compromised or overbroad access.
	8 — Audit Log Management	AI-assisted abuse needs rapid detection through strong logging and review.
	6 — Access Control Management	Least-privilege access is the main limiter on AI-amplified compromise impact.
Recommendation — Tighten account lifecycle controls and remove unnecessary access promptly. Centralise and review logs to detect suspicious AI-enabled activity sooner. Reduce standing access so successful abuse has less room to spread.
NIST Zero Trust (SP 800-207)	SC-1 — Policy Decision Point	AI-era verification benefits from explicit policy decisions before access is granted.
Recommendation — Use policy decisions to verify and bound sensitive requests before execution.

Practitioner Guidance

What to prioritise: Start with controls that reduce spoofability and limit blast radius. If a workflow can be meaningfully abused through impersonation, prompt-based manipulation, or stolen access material, harden the verification step before adding more AI to the process.

What to verify: Confirm that your detection and response teams can independently validate AI-generated findings, because model output should accelerate investigation, not become the evidence base by itself. The practical test is whether a human can still prove who acted, what was accessed, and which trust boundary was crossed.

Common mistake: Treating AI as either a pure productivity tool or a pure threat. In practice, the organisations that do best are those that use AI to compress defender response time while simultaneously tightening identity, approval, and exception handling around high-impact actions.

Practitioner takeaway: The strategic goal is not to outrun AI with more volume, but to make trusted actions harder to fake, easier to verify, and safer to contain when automation is abused.