The conversation on artificial intelligence governance is entering a new phase.
A new bipartisan proposal in the United States would require developers of the largest and most capable AI models to maintain the technical ability to throttle, suspend, or shut down systems should an exceptional risk emerge.
The proposed legislation (known as the AI Kill Switch Act) also introduces incident reporting requirements for the developers it covers and grants federal authorities the ability to order emergency intervention.
Regardless of where organizations stand on the proposal itself, policymakers, technology companies, and cybersecurity leaders all recognize that AI systems require stronger governance as they become more autonomous.
For defenders, however, the discussion raises a different question: If an AI-powered attack can operate at machine speed, how do you stop it inside your own environment?
The answer is to prevent an attack from progressing in the first place.
AI governance extends beyond the model itself
Much of the discussion surrounding AI safety focuses on controlling the model. Questions such as who can deploy it, what tasks it should perform, and how it should be stopped if something goes wrong are becoming central topics in both industry and government. The proposed legislation reflects that broader shift toward requiring technical safeguards alongside operational oversight.
Those are important conversations. But for enterprise defenders, the more immediate concern is what happens after an attacker begins using AI against your organization.
Whether an attack is directed by a human operator or an autonomous system, it still depends on interacting with the target environment. The attacker must execute code, access systems, use credentials, move between devices, and reach sensitive data.
None of those actions can be stopped by a kill switch the defender does not hold. A policy the defender enforces can stop every one of them.
Machine-speed attacks reduce the time available to respond
Cyberattacks often contain pauses. An attacker may stop to analyze results, wait for new instructions, or hand work over to another operator. Those gaps create opportunities for defenders to investigate alerts, validate suspicious activity, and respond before additional damage occurs.
Autonomous AI systems reduce many of those time gaps. They can continuously search for new opportunities, retry failed approaches without fatigue, and immediately move to the next objective when one succeeds.
Detection, investigation, escalation, and response remain important, but every stage has less time to interrupt an attack before it progresses, which makes prevention significantly more valuable.
Zero Trust prevents speed from becoming an advantage
An autonomous attacker only benefits from speed if every successful action unlocks the next stage of the attack.
- Execution leads to privilege escalation.
- Privilege escalation enables lateral movement.
- Lateral movement creates access to additional credentials and sensitive systems.
If any one of those actions is denied, the attack chain stops regardless of speed. That is the principle behind Zero Trust cybersecurity. Rather than attempting to predict every possible attack, Zero Trust defines what is allowed and blocks everything else by default.
Prevention controls do not become less effective simply because requests arrive faster.
Every AI-powered attack still depends on permission
Even the most capable AI cannot compromise systems through reasoning alone. It will still need favorable policy decisions.
To achieve its objective, it still depends on software being permitted to execute, trusted applications being able to perform unintended actions, privileged accounts being available, and network paths remaining open.
That is why the same Zero Trust controls that prevent conventional cyberattacks remain effective against AI-assisted attacks.
The speed and persistence might change, but the technical requirements needed for an attack to succeed remain.
AI governance and Zero Trust cybersecurity fall under the same umbrella
Whether organizations are discussing model evaluations, emergency shutdown capabilities, deployment controls, or oversight processes, the objective is ultimately the same: reducing unnecessary risk while enabling responsible innovation.
Organizations should establish clear policies governing how AI tools are used, what systems they can access, and which data they are permitted to process. At the same time, those environments should be designed so that even if an attacker leverages AI to accelerate an intrusion, they cannot freely execute code, elevate privileges, or move throughout the network.
Governance establishes accountability, while Zero Trust enforces technical boundaries. Both are necessary as AI capabilities continue to evolve.
Legal scholars examining agentic AI frameworks have made a similar point. Reviewing one such framework earlier this year, Eran Kahana of Stanford's CodeX blog wrote that risk identification without control specificity “describes the fire without providing the extinguisher.” For enterprise defenders, the extinguisher is enforcement: controls that evaluate every action against an approved boundary before it executes.
How ThreatLocker strengthens resilience while AI boosts efficiency
Stopping technology from bringing new capabilities and efficiencies to our daily lives should not be the goal here. Instead, the focus should be on defining the guardrails to control it while achieving those objectives safely.
Organizations cannot assume that future AI-powered attacks will provide the same response windows that defenders have relied upon for years. Instead, they should focus on reducing the opportunities available to any attacker before malicious activity begins.
ThreatLocker helps organizations achieve that through a prevention-first, deny-by-default approach. In practice, that is a kill switch of a different kind. It does not wait for an exceptional risk to emerge or an order to be issued. It is applied to every action, before execution, by default.
As AI becomes faster, more autonomous, and more widely adopted, the fundamentals of cybersecurity remain unchanged. The organizations best prepared for the future will be those that prevent familiar attacker actions before speed becomes an advantage.
ABOUT THE AUTHOR
Andrea Pomaranski is a Special Projects IT Engineer at ThreatLocker, working on R&D‑focused technical initiatives across enterprise systems, identity infrastructure, and platform integrations. She studied Computer Science at the University of Florida and began her career in IT before moving into software development, supporting mission‑critical systems in highly regulated production environments. A longtime CTF competitor, Andrea honed her security skills through university, independent study, and past workplaces, including as an active member of UF’s Security and Information Technology (SIT) organization. Outside of work, she maintains a hands‑on homelab where she experiments with hardware and operating systems. She brings a builder’s mindset and a deep interest in how vulnerabilities emerge in real‑world systems to product research and technical investigation.
