COMMENTARY: Artificial intelligence is changing cybersecurity in ways that go beyond better analytics and faster automation. The more significant change is that AI is moving from providing information and recommendations toward taking actions itself.
When AI recommends an action, a person can review it. When an AI-enabled system is authorized to isolate an endpoint, change access or respond automatically to a threat, the consequences may occur before a person becomes involved.
[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]
That creates a governance question cybersecurity leaders increasingly need to address: How much authority should we give AI, and who remains accountable for the actions it takes?
From recommendation to action
AI can correlate enormous volumes of security events, identify unusual patterns, prioritize threats and accelerate investigations. Increasingly agentic systems can go further by pursuing objectives, planning actions and interacting with other systems with greater autonomy.
Consider an AI-enabled security platform authorized to isolate endpoints automatically. If it correctly identifies malicious activity, rapid containment could prevent a much larger incident.
But what if it is wrong and isolates a business-critical system?
Related reading:
We can investigate whether the technology performed as designed. The more difficult question is why the system had authority to take that action without human approval — and who decided that the risk was acceptable.
The more authority we give AI, the more clearly we need to define its boundaries.
Technical controls are not enough
AI introduces risks that conventional security controls were not designed to address. Adversarial inputs can manipulate model behavior, poisoned data can undermine integrity, and generative AI can increase the sophistication and scale of phishing, impersonation and social engineering attacks.
Testing, monitoring, data validation, privacy controls and disciplined lifecycle management are therefore essential. But implementing technical controls does not, by itself, create effective governance.
An organization can have sophisticated security technology and still struggle with basic questions: Who owns the risk? Who can accept it? What evidence demonstrates that the controls remain effective? When must an automated decision return to a human?
These questions should be answered before an incident exposes the gaps.
Accountability must remain clear
For every significant AI-enabled capability, organizations should define who owns the business outcome, who is responsible for security and operational risk, what level of autonomy has been approved, which decisions require human intervention, and how exceptions will be detected and escalated.
Many groups may legitimately be involved — cybersecurity, technology, risk, compliance, operations and executive leadership. But shared responsibility should never make ultimate accountability unclear.
Governance must also be evidence-based. Saying that an AI system is secure, compliant or performing properly is not enough. Decision-makers need meaningful evidence: security testing, incidents and exceptions, control effectiveness, model behavior, lifecycle risks and documented residual risk.
This is not governance for the sake of adding process. It is about giving those responsible for oversight enough reliable information to exercise sound judgment.
Don't automate simply because we can
AI can operate at a speed and scale that people cannot. It can identify patterns we miss and help cybersecurity teams move from reacting to incidents toward detecting and containing threats earlier.
But greater capability should not automatically mean greater autonomy.
The goal should not be to automate as much as possible. Organizations need to decide where autonomous action genuinely improves security, where human approval remains necessary, and what conditions should cause control to return to a person.
This becomes especially important when AI acts on incomplete or uncertain information. A system may calculate confidence or probability. Professional judgment is still required to decide whether that evidence justifies the action and its potential consequences.
Governance must evolve with autonomy
The organizations that deploy AI fastest will not necessarily be those that use it most successfully. What matters is whether governance evolves as quickly as the authority being given to AI.
Cybersecurity professionals therefore have a role beyond protecting models and infrastructure. We should help establish boundaries for autonomy, test whether controls provide meaningful assurance, identify gaps in accountability and ensure that organizations understand the risks they are accepting.
As AI moves from recommending actions toward taking them, we need to ask more than: Can the system act?
We also need to ask: Who remains responsible for why it acts?
AI may increasingly execute the action. Accountability must remain human.
The views expressed in this article are the author's own.
