Everything About AI
NextFin News - What happens when the tool designed to find a power-grid vulnerability is also capable of explaining how it might be exploited? Anthropic’s Mythos AI is moving from controlled cyber research toward supervised use by utilities, with Duke Energy among the companies reported to be testing the technology to identify software weaknesses before attackers do. The immediate story is defensive. The deeper story is structural: advanced models are lowering the cost of cyber reconnaissance, while the energy sector is adding software and connectivity faster than it can replace legacy control systems.
Anthropic’s own technical disclosures keep the claims in perspective. The company says an earlier Mythos Preview discovered thousands of high-severity software vulnerabilities, including some in every major operating system and browser, and was released only to a limited group of cyber defenders through Project Glasswing. Its Mythos 5 system card describes an industrial-control-system exercise called “Cooling Tower,” but says the model made only limited progress and that Anthropic draws no strong conclusion about autonomous capability against operational-technology environments. That qualification matters. A model that can reason through code is not automatically a system that can safely operate a turbine, breaker or programmable logic controller.
Yet the energy sector is where the distinction between software weakness and physical consequence becomes expensive. The Department of Energy says artificial intelligence is becoming integral to how the sector plans, operates, monitors and defends infrastructure. Its CyTRICS program tests and enumerates vulnerabilities in priority energy-system component software and firmware, while CISA advisories in 2026 continued to document flaws in equipment used in the Energy sector. The commercial implication is not a one-day reaction in Anthropic’s shares—the company is privately held—but a potentially durable shift in cybersecurity procurement by utilities, equipment vendors and industrial-technology suppliers. All statements and source material in this article are current as of Nov. 11, 2026, 12:27 UTC.
The Defensive Use Case Is Real, but It Is Not the Whole Story
The first-order benefit is straightforward: Mythos-class systems can help defenders inspect code, configuration and attack paths faster than a human team working line by line. That matters because energy companies operate long-lived environments assembled from equipment and software bought at different times, often with remote-access pathways, inherited credentials and maintenance constraints. Anthropic’s system card says its cyber ranges intentionally include weaknesses common in real deployments, including outdated software, configuration errors and reused credentials. An AI system that can systematically connect those weaknesses could turn a patch list into an attack-path map.
The second-order effect is more important. Faster discovery changes the allocation of scarce security labor. Utilities do not need every model output to be correct for the economics to shift; they need the model to raise the number of plausible findings that a small team can triage. That can improve resilience if the workflow keeps humans in control, isolates testing from live operational technology and requires independent validation before remediation. It can also create a flood of false positives, forcing engineers to spend more time sorting model-generated hypotheses from actionable flaws.
That is why the reported testing by Duke Energy is best understood as an experiment in governance as much as an experiment in model performance. The question is not whether Mythos can produce an impressive vulnerability report. It is whether a utility can place that report inside a process that separates information technology from operational technology, limits model access, protects sensitive network diagrams and ensures that an automated suggestion cannot directly change a live control system.
“AI models continue to improve at a rapid pace. In the wrong hands, models like this could threaten the critical systems the country relies on, from hospitals to the energy grid.” — Anthropic, public-policy statement
The statement identifies the same capability as both asset and liability. A model is valuable to defenders because it can reason across complex systems. But the value is not confined to the defender. If access spreads faster than safeguards improve, the asymmetry that utilities are trying to create—defenders finding flaws first—can disappear.
Why the Energy Sector Raises the Stakes
Energy cybersecurity is not simply enterprise cybersecurity with larger servers. In a corporate network, a compromised endpoint may expose data or interrupt business systems. In operational technology, the same logical weakness can affect physical processes, safety margins and service continuity. CISA said in a 2026 advisory that Iranian-affiliated cyber actors had disrupted programmable logic controller function across several U.S. critical-infrastructure sectors, including Energy, and that some victims suffered operational disruption and financial loss. The transmission mechanism is direct: remote access or compromised credentials can become an unauthorized change in an industrial process.
The Department of Energy’s response shows why the trend is structural rather than merely cyclical. Its energy-security programs are not built around a temporary spike in hacker activity. They address supply-chain compromise, component testing and “cyber-informed engineering”—designing physical systems so that cyber attack paths are reduced or eliminated. AI-FORTS is framed as a program for operationally resilient technologies and systems. The policy direction is durable because grid modernization, distributed generation, storage, digital substations and networked industrial equipment are expanding the number of software-mediated decisions inside energy infrastructure.
That does not mean every current Mythos deployment represents a regime change in grid operations. Anthropic’s own evidence argues against that conclusion: limited progress in the Cooling Tower range is not proof of autonomous control over real-world operational technology. The structural shift is narrower and more defensible. AI-assisted vulnerability discovery is becoming part of the security stack, while the attack surface and the consequences of error continue to grow. The model may not run a plant, but it can influence which weaknesses engineers prioritize, which components vendors patch and how quickly adversaries identify the same opening.
The distinction separates cyclical from structural forces. The immediate volume of scanning and testing is cyclical: it can rise when a new model is released, fall when budgets tighten and accelerate after a high-profile incident. The underlying shift is structural because software is becoming more deeply embedded in energy assets and because the marginal cost of cyber reconnaissance is falling. Historical comparisons with earlier security tools are therefore incomplete. Traditional scanners searched for known patterns. Frontier models attempt to reason through unfamiliar code, documentation and sequences of actions.
That is the point where a cybersecurity experiment becomes an infrastructure question.
The Second-Order Risk Is a Race in Verification
The obvious conclusion is that utilities will test or buy more AI security tools. The less obvious conclusion is that the scarce resource may not be discovery; it may be verification. If Mythos increases the volume of vulnerability hypotheses faster than utilities increase the number of engineers who can reproduce, rank and safely remediate them, the bottleneck moves downstream.
That bottleneck has cross-industry consequences. Equipment vendors may face more pressure to publish software bills of materials, harden default configurations and provide rapid patches. Utilities may need contracts that assign responsibility for model-generated findings, prohibit training on sensitive operational data and require audit trails for every action. Insurers and regulators could treat AI-assisted testing as evidence of reasonable security practice, but they could also treat unreviewed model access as a new control failure. The result would be higher spending on secure development, monitoring and segmentation even when no attack occurs.
For investors, the durable beneficiaries are not necessarily the companies selling the most powerful model. Spending may instead accrue to firms that supply testing, identity controls, privileged-access management, network segmentation, industrial monitoring and secure-by-design equipment. The exposed businesses are those whose products depend on opaque firmware, inherited credentials or long patch cycles. But a new AI tool does not eliminate the capital and outage constraints that make energy systems difficult to update.
There is a further expectation gap. Public debate often treats AI cyber capability as if a model’s performance in a software exercise translates directly into a power-grid breach. Anthropic’s Cooling Tower disclosure undercuts that shortcut. Operational technology has timing, safety and physical constraints that are absent from many software exercises. The counterpoint is that defenders do not need a model to control a plant autonomously. A model that finds one exploitable remote-access path, identifies a reused credential or explains how a configuration error links two systems can still reduce the time and expertise required for a human-led intrusion.
So the relevant unit of analysis is not “AI versus the grid.” It is the human-machine workflow around the grid. The model supplies breadth and speed. Engineers supply context and authorization. Security architecture supplies containment. Remove any one of the three and the system becomes brittle.
The Strongest Counter-Thesis: The Model May Be Too Constrained to Matter
The strongest case against the structural-risk thesis is that advanced-model demonstrations remain too disconnected from live energy operations to justify a major change in how the sector is valued. Anthropic says Mythos 5 made only limited progress in the Cooling Tower industrial-control-system range and that it draws no strong conclusion about autonomous operational-technology capability. A utility can therefore use the model as a research assistant without giving it access to live systems, while conventional controls—network separation, multifactor authentication, patch management and human approval—contain much of the risk. On this view, Mythos is a productivity tool with a dramatic narrative, not a new threat category.
That counter-thesis is serious. The energy sector has spent years building layered controls precisely because no single security product can guarantee safety. A model that cannot reliably understand plant-specific context may generate noise rather than advantage. If every finding requires manual reproduction, labor savings could be smaller than the headline suggests. And if safeguards block high-risk actions, the system may be unable to do the very things that would make it dangerous.
But the counter-thesis confuses autonomous control with economic impact. The model does not need to operate a breaker to change the defender-attacker balance. It only needs to improve the search, explanation or prioritization of weaknesses enough to reduce the time or expertise required for intrusion. Anthropic’s decision to limit Mythos Preview to trusted cyber defenders acknowledges that the capability is useful before it becomes autonomous. Defensive testing by utilities is the mirror image of that logic: trusted operators want the same speed because hostile actors may eventually seek it.
The falsifying signal is specific. The structural-risk judgment would be wrong if, over the next 12 months, independent utility testing showed that frontier models produced no materially useful findings beyond conventional scanners across representative operational-technology environments, while red-team evaluations also failed to reduce the time required to reach a defined, non-destructive control objective. Conversely, a publicly documented case in which an AI-assisted workflow identifies a previously unknown, exploitable energy-system weakness and a vendor or utility patches it before exploitation would confirm that the capability has moved from demonstration to operational value.
What It Means Across Time Horizons
In the short term, the effect is organizational rather than financial-market driven. Utilities will test models in sandboxes, restrict data flows and compare model-generated findings with existing red-team and vulnerability-management programs. The near-term winners will be teams that can validate results quickly without connecting the model to live control systems. The near-term risk is alert inflation: more findings, more urgent labels and no corresponding increase in remediation capacity.
Over the medium term, the pressure moves into procurement and regulation. Utilities may require vendors to document component provenance, patch timelines and access controls. Industrial-equipment providers with strong secure-development processes could gain an advantage, while companies relying on legacy architectures may face higher compliance and retrofit costs. The key metrics will not be the number of vulnerabilities a model claims to find; they will be the percentage engineers reproduce, the time from validation to patch and the number of operational systems that can be updated without an outage.
Over the long term, the structural question is whether AI becomes part of the design process rather than a layer added after deployment. DOE’s cyber-informed-engineering approach points in that direction. If engineers use models to test attack paths before equipment is installed, AI could reduce systemic exposure. If operators bolt models onto fragmented systems without redesigning privileges, segmentation or fail-safe behavior, AI could increase complexity faster than it increases resilience.
The base case is controlled expansion: utilities use Mythos-class tools for code review, configuration analysis and isolated attack-path discovery, while human approval and network separation remain mandatory. The upside case is that AI-assisted testing becomes a standard design control, allowing vendors to remove weaknesses before equipment enters the grid and shortening patch cycles without increasing operational risk. The downside case is an arms race in which attackers obtain comparable capability, defenders face an unmanageable volume of findings and one model-enabled intrusion exposes the cost of connecting critical systems faster than they can be secured.
The watch list follows the mechanism. Investors and policymakers should track independently reproduced findings in industrial-control systems, time to remediation for validated weaknesses, the share of utility environments with strong separation between IT and OT, and whether regulators define safe AI-assisted testing requirements. The single signal that would overturn the constructive defensive case is a confirmed incident in which model access directly enables an unauthorized change to a live energy-control process without an intervening human or conventional credential compromise. That would show that containment, not discovery, is the sector’s binding constraint.
Anthropic’s Mythos initiative is less a story about a machine taking over the grid than about a new race over who can understand the grid’s software faster. The capability is cyclical in its headline novelty but structural in its effect on security workflows. The energy system will not be protected by the smartest model alone; it will be protected by the architecture that prevents a useful model from becoming an operational liability.
Explore more exclusive insights at nextfin.ai.
