AI is changing cybersecurity: Identity is your best defense

Artificial intelligence is reshaping cybersecurity on both sides of the fight. Security teams use it to automate threat detection, accelerate investigations, and cut alert fatigue. Attackers use the same technology to run faster, more convincing, and more scalable campaigns.

The question is no longer whether AI affects your security strategy. It already does. The real question is whether your foundations can withstand attacks that now unfold in minutes rather than days, and whether you know what a stolen credential is actually allowed to do once it works.

That second question is an identity question. It is also the one most organizations cannot answer quickly.

AI is accelerating known attacks, not inventing new ones

Most coverage of AI in cybersecurity focuses on novel attack techniques. The measured data points somewhere less dramatic and more actionable: AI is making familiar attacks faster and cheaper, not fundamentally different.

Verizon’s 2026 Data Breach Investigations Report analyzed AI-assisted attacker activity in collaboration with Anthropic, covering 793 threat actors sanctioned for policy violations between March 2025 and February 2026. In the median case, those actors sought AI assistance across roughly 15 distinct MITRE ATT&CK techniques, and fewer than 2.5% of the techniques observed were classified as rare. Phishing-related activity accounted for 44% of AI-assisted initial access.

In other words, attackers are using AI as an operational tool to scale techniques defenders already know, not to unlock unknown ones.

That matters for how you prioritize. Tasks that once took hours of manual preparation now take seconds: personalized phishing emails that mirror a colleague’s writing style, multilingual social engineering campaigns, target research assembled from public sources, and voice or video deepfakes used to impersonate executives. The barrier to running an effective campaign has never been lower, and the window to detect a compromised account before lateral movement begins keeps shrinking.

The techniques are recognizable. The timeline is not.

On AI-generated phishing: How AI-powered phishing has changed what a suspicious email looks like

Attackers break in through vulnerabilities, then move with credentials

This is where the 2026 data gets interesting, and where a lot of identity-security messaging is now out of date.

For the first time in the DBIR’s 19-year history, exploitation of vulnerabilities became the most common initial access vector, accounting for 31% of breaches, up from 20% the previous year. Credential abuse as a standalone initial vector fell to 13%. Read only the headline and you would conclude that identity threats are receding.

The full breach data says otherwise. When Verizon counts credential abuse anywhere in the breach progression rather than only at the first step, it appears in 39% of breaches, more than any other single technique in the dataset. Stolen credentials hold steady at 36% across breach action varieties. In System Intrusion breaches, stolen credentials and exploited vulnerabilities are equally common at 39% each.

The pattern is a two-stage one:

1. Entry increasingly comes from an unpatched application, a compromised vendor, an exposed remote service, or a malicious package.
2. Movement still comes from credentials: passwords, API keys, OAuth tokens, service account credentials, cloud keys, deployment tokens, and any standing permission attached to them.

This distinction has a practical consequence that vulnerability management alone cannot address. You can patch the flaw, remove the malicious package, and rebuild the host, and still leave a copied API key or service account password in circulation. That credential keeps working until someone finds and revokes it.

Microsoft’s Digital Defense Report 2025 reinforces the same point from a different dataset: identity-based attacks rose 32% in the first half of 2025, and more than 97% of them were large-scale password attacks (password spray and brute force against predictable, reused credentials).

So identity isn’t usually how attackers first get in. In modern cybersecurity, they often enter through vulnerabilities or exposed services, but identity is what decides how far they can go, how quickly they can move, and what the breach ultimately costs.

That’s because the “perimeter” is no longer a fixed network boundary. Cloud services, remote users, SaaS applications, and machine workloads operate across changing locations and dynamic environments. As a result, identity and access controls become the consistent enforcement layer for every request, whether it comes from a laptop, a browser session, an API call, or a workload credential.

Patching reduces the chance the door opens. Identity governance determines what’s behind it: which systems an attacker can reach, which privileges they can abuse, and which risks you can contain before lateral movement accelerates.

On identity as the perimeter: Why identity is the new perimeter, and how SMBs can defend it with limited resources

Non-human identities are the fastest-growing blind spot

A non-human identity (also called a machine identity or workload identity) is any credential-holding entity that isn’t a person: service accounts, applications, workloads, scripts, CI/CD pipelines, bots, and increasingly AI agents. In AI security, these identities are what let agents authenticate to systems, retrieve data, call APIs, and perform authorized actions as part of business processes, without a human logging in.

The risks come from how these identities are typically managed: they often hold elevated permissions, may not have strong authentication controls, and can be hard to track or revoke quickly. When an AI agent’s credentials are shared, over-privileged, or poorly governed, a compromise can turn into fast, scalable access, because the agent can keep acting with the permissions it was granted.

They now outnumber human identities in most environments, and they are governed far less rigorously. Microsoft’s 2025 report notes that attackers are deliberately shifting attention toward service accounts and their credentials, precisely because those accounts tend to combine elevated privileges with weak security controls. A service account rarely has MFA, often has a password that has not rotated in years, frequently has more permissions than its function requires, and usually has no clear owner who would notice it behaving strangely.

Four controls close most of this gap:

1. Inventory them. You cannot govern what you have not enumerated. Include service accounts, break-glass accounts, API keys, and agent credentials, and assign a named owner to each one.
2. Scope their permissions. A service account should hold the narrowest set of rights its function requires, on the narrowest set of systems.
3. Manage their secrets centrally. Vault the credentials, rotate them on a schedule, and eliminate shared or hardcoded secrets.
4. Make their activity auditable. If a workload identity starts authenticating from a new location or reaching systems it has never touched, someone should see it.

Ungoverned machine identities are how attackers move through an environment faster than a team can respond, and they are the category most likely to be missed in an access review built around employees.

AI is helping defenders too, but adoption is the bottleneck

Security teams are getting real value from AI. Modern platforms correlate millions of events in real time, surface anomalies, cluster related alerts, and summarize complex incidents in seconds. Practically, that means analysts spend less time triaging and more time on the incidents that matter: prioritizing high-risk alerts, accelerating response, recommending remediation, and automating repetitive operational work.

The gap is deployment, not capability. Devolutions’ State of IT Security in SMBs survey found that 71% of SMBs plan to increase their use of AI in cybersecurity, but only 25% use it today and 40% have not started at all. Meanwhile, fewer than a third of respondents said they could identify an attack quickly enough to limit the damage.

That is the uncomfortable pairing: attackers have already operationalized AI, and most defenders have it on a roadmap.

Worth being clear about what AI can and cannot fix, though. It compresses detection and investigation time. It does not compensate for excessive privileges, unmanaged secrets, or an admin account nobody remembers creating. An AI-assisted SOC that detects a compromised account in 4 minutes instead of 40 has bought you real time, but only if that account’s permissions are narrow enough that the time is worth something.

On the AI arms race: Offensive AI vs. defensive AI: who has the upper hand in 2026

The fundamentals decide your blast radius

As AI dominates the conversation, it is tempting to believe the next generation of tools will solve the underlying problem. It won’t, for a simple reason: the controls that limit damage after a compromise are the same ones they have always been. What changed is how quickly a gap gets found and exploited.

Organizations that consistently reduce risk work through a recognizable sequence:

• Enforcing the principle of least privilege, so each account holds only the rights its role requires
• Securing privileged accounts with privileged access management (PAM)
• Deploying phishing-resistant multi-factor authentication on privileged access and critical systems
• Managing secrets and service accounts centrally, with rotation
• Reviewing and certifying user access on a defined cadence
• Monitoring privileged activity continuously
• Implementing just-in-time (JIT) privileged access wherever standing access can be removed

None of these are new. Two of them are unusually high-leverage right now.

The first is phishing-resistant MFA: strong, cryptographic authentication that uses passkeys, FIDO2 security keys, or certificate-based methods to bind login to the legitimate service and resist credential relay. In practice, it means an attacker who steals a password still can’t complete authentication, because the second factor can’t be replayed or proxied the way traditional one-time codes can.

Use it first for privileged accounts and critical systems, and extend it to remote access and other high-risk entry points where stolen credentials are most likely to be used. This is one of the highest-leverage defenses in cybersecurity because it directly reduces the risks created by credential theft: it breaks the path from “stolen” to “usable,” limiting the threats that identity compromise can turn into real breach activity.

The second is reducing standing privilege. Least privilege limits what any user, service account, or workload can do; role-based access control (RBAC) assigns permissions by job function rather than by request history. Together they cap the number of actions an attacker can take with a working credential, constrain lateral movement, and contain the blast radius. The same survey found that 52% of SMBs still manage privileged access manually (spreadsheets, shared vaults, or no formal system at all) and only 32% use a dedicated PAM tool. Manual processes introduce error, obscure visibility, and delay revocation exactly when speed matters most.

When attackers automate reconnaissance and privilege escalation, every unnecessary permission, dormant account, and unmanaged secret becomes an opportunity that gets found sooner than it used to.

On elevated access: How PIM and PAM together turn privilege elevation into short, auditable, governed events

Conditional access turns identity into a real-time control

Authentication at login is a single check at a single moment. Conditional access and continuous verification turn identity into an ongoing control.

Conditional access evaluates signals (user identity, device posture, location, behavior, and the sensitivity of the resource requested) and decides whether to allow, restrict, or block the request. Continuous verification then re-evaluates those signals during the session. If risk rises mid-session because a device falls out of compliance, a location pattern looks impossible, or behavior deviates from baseline, the system can step up authentication, reduce permissions, or terminate the session.

The value is timing. These controls shorten the interval between a credential being stolen and that credential stopping working, which is the interval AI-accelerated attacks are designed to exploit.

Visibility is what you’re actually missing

The most common question after an incident is disarmingly simple: why did this account still have access?

The answer is usually visibility, not technology. Most organizations own tools capable of answering the question. Few can answer it in minutes.

Try these five on your own environment:

• Who currently holds administrative privileges?
• Which privileged accounts have gone unused for 90 days or more?
• Which temporary access grants were never revoked?
• Where are service account credentials stored, and when did they last rotate?
• Which identities can reach your most critical systems today?

If any of those takes hours or days to answer, an attacker working from a stolen credential has more information about your environment than you do. Strong identity governance is what converts those questions into queries rather than investigations.

On what’s ahead: Cybersecurity trends in 2026: identity, AI, and incident response

What matters isn’t how much AI you have

AI will keep evolving on both sides. Attackers will find new applications for it; defenders will keep improving detection, automation, and response. That arms race isn’t slowing.

But the organizations that come through it well won’t be the ones with the most AI. They’ll be the ones with the tightest control over identities, privileges, and access, because that control is what determines the cost of a breach that has already started.

At Devolutions, we treat identity security as a foundational pillar rather than a feature. AI can help you detect faster and operate more efficiently. It cannot substitute for disciplined identity governance or strong privileged access controls.

When an incident happens, the important question isn’t whether AI was involved. It’s what that compromised identity was allowed to do.

See the data: The Devolutions State of IT Security in SMBs survey report: PAM adoption, AI readiness, and incident response gaps

Frequently asked questions

These short answers recap the identity questions that matter most when AI is accelerating familiar attacks: how they start, how they spread, and which controls actually limit the damage.

Is AI creating new cyberattack techniques, or just faster versions of existing ones?

Primarily faster versions. Verizon’s 2026 DBIR analysis of 793 sanctioned threat actors found that in the median case, attackers used AI across roughly 15 known MITRE ATT&CK techniques, and fewer than 2.5% of observed techniques were classified as rare. AI’s main effect is to compress preparation time and increase the scale and quality of familiar attacks, particularly phishing, which accounted for 44% of AI-assisted initial access, rather than to introduce novel attack methods.

If vulnerability exploitation is now the top initial access vector, why does identity security still matter?

Because initial access is only the first step. The 2026 DBIR found that exploitation of vulnerabilities accounts for 31% of breach entry points, but credential abuse appears in 39% of breaches when counted anywhere in the attack progression, more than any other technique. Attackers commonly enter through an unpatched system, then use stolen credentials to move laterally and escalate privileges. Patching controls whether an attacker gets in; identity and privileged access controls determine how far they get and how much damage they cause.

What is a non-human identity, and why is it a security risk?

A non-human identity is any credential-holding entity that isn’t a person: service accounts, applications, workloads, scripts, CI/CD pipelines, bots, and AI agents. They are a risk because they typically combine elevated privileges with weak controls: no MFA, rarely rotated passwords, broader permissions than required, and no clear owner monitoring their behavior. Microsoft’s 2025 Digital Defense Report notes attackers increasingly target service accounts for exactly this reason. Governing them requires an inventory with named owners, scoped permissions, centrally vaulted and rotated secrets, and auditable activity logs.

Why does least privilege matter more against AI-driven attacks?

Least privilege limits what an attacker can do with a credential that already works, and AI shortens the time between credential theft and privilege escalation. When reconnaissance and escalation are automated, every unnecessary permission is found and used faster than before. Least privilege and role-based access control cap the number of actions available after a compromise, restrict lateral movement, and contain the blast radius, which is what determines the cost of an incident once detection has already been outpaced.

What is the single most effective identity control to deploy first?

Phishing-resistant multi-factor authentication on privileged accounts and critical systems. Microsoft reports that more than 97% of identity attacks are large-scale password attacks, and that phishing-resistant MFA prevents up to 99% of them. After MFA, the highest-leverage step is removing standing privileged access in favor of just-in-time elevation, which shrinks the window during which any stolen privileged credential is usable.