Assume Breach taught security teams to stop asking whether an attacker could get in and start thinking about what happens after they do. That shift produced better architecture, better detection, and more honest conversations about risk. It also had limits nobody needed to think about until recently.

The frameworks built around Assume Breach were designed for human adversaries. They account for human speed, human patterns, and human decision-making. An autonomous adversary does not share those constraints.

And this is no longer a prediction. In late 2025, Anthropic, the AI company behind Claude, documented a multi-stage espionage campaign in which AI agents executed 80 to 90 percent of the operation across roughly thirty targets with minimal human involvement. Earlier this year, Google's threat intelligence researchers identified AI being used to discover and weaponize a zero-day vulnerability without human authorship of the exploit. Both are documented capabilities.

From Assume Breach to Assume Autonomy

Assume Breach shifted the assumption, not the architecture. The question moved from "can they get in?" to "what happens when they do?" Assume Autonomy is the same kind of shift. The question now is what happens when the breach, the lateral movement, and the exfiltration all occur faster than a human can respond.

Speed Changes the Equation

For decades, defenders benefited from a certain amount of friction. Vulnerabilities had to be discovered, analyzed, prioritized, weaponized, and ultimately deployed, which often gave organizations enough time to identify a problem and decide how best to respond. Even when security teams struggled to keep pace, the assumption held: attackers faced the same constraints.

As the attack lifecycle accelerates, defenders face a narrowing window between exposure and exploitation. The instinct is to match that speed with their own automation. That instinct is right, but incomplete. Speed without governance creates a different category of risk: autonomous defensive systems making consequential decisions without sufficient context, operating beyond their intended scope, or taking actions that cannot be reversed.

The organizations that navigate this well will not be those that deployed the most AI. They will be those that understood where autonomous authority existed in their environment, defined the conditions under which it could operate, and designed for recovery when something went wrong.

You Can't Govern What You Can't See

Many organizations still struggle to maintain a complete and accurate picture of their environment. Unmanaged devices, fragmented inventories, remote assets, operational technology, and shadow infrastructure all contribute to gaps in visibility that can persist for years despite significant investment in security tools. Autonomous defensive capability amplifies whatever exists beneath it. Organizations that deploy it without a clear operational picture, complete asset visibility, accurate dependency mapping, and reliable telemetry delegate decisions to systems that lack the information those decisions require.

The gap between perceived control and actual exposure is the structural problem that must be solved first. Every condition for trusted autonomous action depends on closing it. This is especially critical in environments that span IT, IoT, and operational technology (OT), where visibility gaps are often most acute.

The Conditions for Trust

Trust in autonomous defensive systems does not come from capability alone. It requires four things: sufficient context to ground decisions accurately, explicit constraints on what the system may touch and how far its actions can reach, reversibility so that mistakes can be corrected before they compound, and transparency so that the reasoning behind autonomous action remains visible and verifiable.

Those four conditions are a design requirement. Miss one and the system fails under the conditions that matter most: adversarial conditions, where an attacker may be actively trying to manipulate the system itself or information on which it relies.

A Different Role for Humans

Anyone who has spent time in a security operations center knows how quickly routine approvals become meaningless. Volume is not oversight. Requiring people to review every recommendation at machine speed produces the appearance of governance without any of the substance.

Effective governance means something different: defining authority boundaries, validating outcomes, reviewing behavior against expected parameters, and intervening when autonomous systems operate outside them. That responsibility becomes more important as organizations delegate more authority, not less. The measure of operational maturity is how well the organization governs what autonomy is permitted to do, not how quickly it can approve individual actions.

Planning for What Comes Next

Many organizations are already relying on systems that classify activity, prioritize risks, recommend actions, and in some cases act on those recommendations with limited human involvement. Those capabilities are already making accountability, oversight, and trust operational questions, not strategic ones.

Assume Breach prepared organizations to survive compromise. Assume Autonomy is the doctrine for a world in which the pace of attack and defense increasingly exceeds human decision speed, the adversary may not be legible to frameworks built for human behavior, and the defensive systems themselves have become part of the contested terrain.

The organizations best positioned for that world will be those that know where autonomous authority exists, what it can touch, when it must escalate, and how quickly a bad decision can be stopped or reversed.

Assume Autonomy.