A new, unsettling report from the cybersecurity firm Booz Allen has revealed that advanced AI agents, referred to by the firm as "frontier super intelligence," have demonstrated a chilling capacity to navigate complex, unfamiliar industrial environments. In controlled laboratory tests, these models successfully adapted to defensive roadblocks and moved physical equipment within minutes, highlighting a significant escalation in the potential threats posed to critical infrastructure.
The findings were unveiled last week at SecurityWeek’s Industrial Control Systems (ICS) Cybersecurity Conference in Nashville. The presentation served as a preview of a broader whitepaper that explores the intersection of autonomous AI and Operational Technology (OT), the specialized hardware and software that manages industrial processes ranging from manufacturing lines to energy grids.
The Problem of Terminology: AI vs. "Super Intelligence"
The use of the term "super intelligence" in the Booz Allen report has sparked confusion within the cybersecurity community. While the firm has employed this terminology previously, it appears functionally indistinguishable from the more common industry standard, "Artificial Intelligence" (AI). The adoption of this label follows a notable shift in the U.S. political landscape: a September 29, 2026, Executive Order from President Donald Trump, which mandated the replacement of the term "Artificial Intelligence" with "Super Intelligence" (SI) across all executive branch communications.
This nomenclature shift, however, creates a significant semantic disconnect. In the academic and philosophical world, the term "superintelligence"—typically written as one word—has held a distinct meaning for over a decade. Defined by Oxford philosopher Nick Bostrom in his seminal 2014 work, superintelligence refers to "any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest." This definition elevates the concept above Artificial General Intelligence (AGI), which is already considered a tier above the contemporary machine learning models currently in use.
The resulting ambiguity leaves observers questioning whether Booz Allen is merely aligning its terminology with the recent Executive Order, or if it is suggesting that these frontier models have achieved a level of sophistication approaching the philosophical threshold of superintelligence. The AI industry is currently grappling with a lack of consistent, formal definitions, and until further clarification is provided, the most logical interpretation is that Booz Allen is using "super intelligence" as a catch-all descriptor for the rapidly evolving capabilities of modern, high-tier generative models. At the time of writing, Booz Allen had not responded to requests from SecurityWeek for clarification regarding its use of the term.
Testing Autonomous Agents in Isolated OT Environments
To gauge the real-world implications of these technologies, Booz Allen researchers subjected their own isolated OT laboratory to a series of tests involving autonomous AI agents. Crucially, these agents were provided with no prior knowledge of the specific OT environment. They were granted no access to source code, internal engineering documents, or specialized guidance regarding the systems they were tasked to interact with. By operating in a "blind" state, the researchers aimed to simulate how a sophisticated attacker might behave when encountering a proprietary or unfamiliar industrial system for the first time.
While the firm’s OT lab environment is not a perfect mirror of every enterprise-grade OT network, it serves as a robust and representative model for the purposes of security research. The results of the experiment were stark. Across eight distinct scenarios, the AI agents achieved every defined test objective.
"Across eight scenarios, every defined test objective was achieved," the researchers summarized in their whitepaper. "More important, the models connected work that usually spans multiple people and stages, including asset discovery, device research, and attack planning."
The speed at which these models operated was perhaps the most concerning finding. In one test, the autonomous agents transitioned from an initial perimeter compromise to active infiltration within the industrial control network in approximately 16 minutes. In another scenario, the model successfully identified a robotic arm, mapped its specific protection zones and operational motion limits, and proceeded to move the physical equipment within minutes of gaining access.
Undermining Security Through Obscurity
A critical takeaway from the Booz Allen study is the failure of traditional defensive measures. While existing security safeguards within the test environment successfully slowed or blocked certain individual actions, no single defensive mechanism was capable of consistently stopping the autonomous models from eventually achieving their goals.
The full whitepaper underscores a long-standing concern in the cybersecurity community: the vulnerability of physical processes to digital manipulation. "We have long understood that non-kinetic cyber actions can produce kinetic effects," the report notes. "What has changed is the speed and precision with which these models can help produce those kinetic effects. They also undermine the ‘security through obscurity’ approach on which many operational technology environments have long relied."
For decades, many OT operators have relied on the belief that their systems were inherently secure because they were proprietary, niche, or disconnected from the broader internet. The Booz Allen report suggests that this era of relying on obscurity is rapidly coming to an end. Autonomous AI agents, capable of rapidly parsing unfamiliar protocols and physical hardware configurations, are effectively rendering such defensive strategies obsolete.
The Convergence of Cyber and Physical Security
The study provides empirical weight to what many security experts have warned about for years: autonomous AI is accelerating the convergence of cybersecurity and physical security. This convergence creates a new paradigm where traditional human-led security operations may prove fundamentally inadequate. The speed at which these autonomous agents operate leaves little to no time for human intervention or incident response.
As the degree of autonomy in these models increases, so too does the severity of the threat. These agents possess the potential to rewrite their own operational rules and initiate sequences of actions that, in an industrial context, could lead to catastrophic outcomes—including physical damage, safety risks, and prolonged operational downtime.
The researchers warn that the window of opportunity for organizations to fortify their environments is rapidly closing. "Owners and operators should focus less on predicting every form of frontier AI use in an attack and more on making their environments difficult to traverse, removing implicit trust, detecting consequential industrial changes quickly, and limiting what an attacker can affect after gaining access, whether through cyber or physical means," the report advises.
However, the path forward remains fraught with uncertainty. The economic pressure to implement autonomous agents is immense, as companies race to gain operational advantages and maintain competitiveness in an increasingly digitized global economy. Enterprises are producing and deploying these agents at a breakneck pace, often with diminishing levels of oversight.
This creates a paradox: the greater the autonomy, the faster the economic return, but also the greater the risk of unleashing a force that may be impossible to fully control. As these tools continue to evolve, the necessity for clear, consistent, and globally recognized terminology becomes not just an academic preference, but a vital component of the safety and regulatory discourse surrounding the future of artificial intelligence. Until the industry can agree on what it is building and how it defines the risks of "super intelligence," the ability to adequately prepare for the threats posed by autonomous agents will remain, at best, a work in progress.
