Static AI guardrails are inherently limited

As organizations adopt generative AI, many still assume that the right set of guardrails will be enough. The problem is you can’t anticipate every way these systems might be misused, abused or attacked. What NIST has done is put a mathematical foundation under that intuition.

In recent research building on Gödel’s incompleteness theorems, which showed that any system built on a fixed set of rules will always have gaps, NIST demonstrates that there is no finite set of guardrails that can be universally robust against adversarial prompts. In plain terms, if your defense is based on a fixed set of rules, there will always be inputs that bypass them. Not because the rules are badly written, but because the problem space is bigger than static rules can ever cover.

This is not new in cybersecurity - detection rules have always had to live with this trade-off. What is different with GenAI is the scale and shape of that problem. These systems are built on human language, and human language is not bounded. It is fluid, contextual and deliberately ambiguous. The number of ways intent can be hidden is effectively limitless. You are not defending against a defined protocol or a fixed exploit chain. You are defending against the entire expressive capacity of people.

So attempting to create a complete set of rules is the wrong starting point. It assumes the problem can be deterministically described. NIST’s work shows that it cannot. Organizations still need a way to manage AI risk, but the traditional approach of defining allowed and disallowed patterns is always going to lag behind what is actually happening. The same input can be benign in one context and risky in another, and static rules struggle to capture that distinction.

The question then is what fills that gap?

AI security must shift from rules to behavior

What's required is a shift in what you are trying to understand. Rules try to describe what should and shouldn't happen. Behavior shows you what is happening. Or to put it another way, if inputs are unbounded and adversaries adapt, the only stable signal is behavior.

In a GenAI context, that means analyzing how an AI model is being used, how prompts evolve over time, how outputs are shaped, and where AI agent interactions start to drift from what is expected. It means moving from static definitions of bad to a more dynamic understanding of intent.

Instead of trying to predict every bad prompt, you focus on identifying when behavior starts to move outside expected norms. Instead of asking whether a single input matches a rule, you ask whether the overall pattern of activity makes sense for the system and how it’s being used.

Guardrails remain important but they are only one layer

This does not eliminate the need for guardrails. They still play a role. But they will never address the entire problem space and are simply one part of your defense in depth approach.

NIST’s proof is useful because it makes this explicit. It removes the assumption that with enough effort, a complete rule set is achievable. It isn’t.

Once you accept that, the shift becomes unavoidable. This is no longer a problem of writing better rules, but of understanding behavior in a space where the possible inputs are effectively unbounded.

For security leaders, that changes the nature of the problem. It is less about defining what should be allowed, and more about recognizing when something is no longer consistent with expected behavior.

That does not remove the need for guardrails, but it does change their role. They set boundaries, but they do not define understanding. The gap between the two is where risk now sits.

In the end, this is what “can’t be solved with rules” really means. Rules will always leave gaps, and those gaps are not theoretical. They show up in how systems actually behave Not what we expect them to do, or what we intended them to do, but what they are doing in practice. That is where the signal is, and increasingly, that is where the security problem sits.

References:

https://www.nist.gov/news-events/news/2026/06/nist-mathematical-proof-supports-transition-continuous-monitor-and-update

https://ieeexplore.ieee.org/document/11475847

The Reality of Securing AI in Motion

Traditional security tools were built for environments defined by fixed rules and predictable workflows. But AI behavior is non-deterministic. The same prompt can produce different outcomes, and risk often emerges gradually as AI behavior adapts, and permissions drift over time. This creates a constantly shifting environment where security teams are working to define control in a system that resists stability. “In AI security, yesterday's priorities can become tomorrow's blind spots. The landscape shifts that fast,” warned the SVP and Head of Technology and Cybersecurity of a real estate investment trust. Conventional approaches, which rely on establishing and maintaining a steady baseline, struggle to keep up with that level of change.

At the same time, AI adoption is accelerating across organizations, often faster than security teams can implement the controls needed to manage it. “The car is being built while it’s already on the road,” explained the CISO of a global private fund administrator. “The threats we're securing against today won't be the threats we're facing tomorrow. What kept us up three months ago looks nothing like what we're dealing with today.”

As businesses move quickly to unlock value from AI, security teams are left closing gaps in real time, while also facing adversaries who are using AI to make their attacks more scalable, adaptive, and difficult to detect. In this recent roundtable discussion of CISOs and security leaders, five themes emerged around AI cyber risk.  

1. AI agents with human access but no human judgment

In Darktrace’s 2026 State of AI Cybersecurity report, 96% of the surveyed security professionals agree that AI significantly improves the speed and efficiency with which they work. Yet, 92% admitted that they’re concerned with the security implications of the use of AI agents across their workforce.

AI agents now operate with human-level permissions across systems, acting at machine speed, orchestrating actions across platforms, and making decisions without the judgment or caution a person would apply. Unlike human users, they cannot be expected to pause and question whether a given action is appropriate.

Their identities are also difficult to inventory, govern, and audit. As agents become easier to deploy than legacy IT systems ever were, organizations are quickly losing track of what is running, what it has access to, and what it is doing. This creates a growing class of highly privileged, autonomous actors operating without the visibility or oversight that traditional identity and access controls were designed to provide.“While AI adoption is critical to running a modern business, AI alone can’t solve all our cybersecurity challenges,” said a global financial sector CISO. “We still need think critically and use human judgement. Those are two things AI can’t do.”

This lack of human judgment becomes especially risky as new architectures, such as Model Context Protocol (MCP), can expand how agents connect to data, tools, and external systems. By design, MCP enables agents to dynamically discover and interact with new resources, increasing flexibility but also introducing new pathways for unintended access, data exposure, or abuse if not properly governed.

The CISO of a fund administrator highlighted one emerging vector as an example: rogue MCP servers. “Our developers want to move quickly and bring value to the business, but technologies like these can unintentionally expose sensitive data in ways that would never have happened before.”

2. Increased digital complexity and expanded attack surface

AI activity rarely stays contained. A single prompt can trigger a chain of actions across networks, email, cloud infrastructure, SaaS platforms, endpoints, identity systems, and development environments, spanning systems that were never designed to be secured as a single, connected flow. This expands both the scale and complexity of what security teams need to monitor and defend.

Yet no single control has visibility across that entire chain. “You can’t defend effectively what you can’t see,” cautioned the private fund administrator CISO. As AI-driven activity moves fluidly across environments, gaps in coverage become inevitable, creating blind spots that attackers can exploit.

Threat actors are already capitalizing on this lack of visibility. “Threat actors have advanced their use of generative AI to launch more convincing phishing campaigns, automate social engineering, and scale attacks with greater precision down to the individual level,” said the SVP of Technology and Cybersecurity for the real estate investment trust. What was once manual and targeted can now be automated and personalized at scale, making attacks harder to detect and easier to execute.

At the same time, the pace of exploitation is accelerating. As a global CISO operating across 40+ countries described it: “Zero-day vulnerabilities are no longer zero day; it’s minus one day. By the time you get to it and address it, it’s already a problem.” By the time risk is identified, it has often already been realized.

The result is a rapidly expanding and increasingly interconnected attack surface that challenges security teams to maintain visibility, context, and control across AI-driven activity.

3. Shadow AI is already everywhere

76% of organizations now cite shadow AI as a problem, one that is spreading through organizations in ways that are hard to track and even harder to control.

Employees are experimenting with publicly available Gen AI tools. Teams are spinning up low-code automations on their own. SaaS providers are quietly embedding AI into existing products. Developers are plugging AI services directly into workflows, often without pausing to consider what that exposure means.

The result is a lack of visibility into:

• What AI tools are being used
• What data those tools can access
• Where prompts and outputs are going
• Which AI agents are interacting with enterprise systems

The SVP of Cybersecurity at a real estate investment trust described the shift: “Before, I was worried about someone sending data erroneously to their personal email. Now we have all these agents online that people are utilizing, and we’re looking at those vectors as well.” For security teams, this means operating without a complete view of how AI is being used, what it can access, and where risk may already be emerging.

4. Built-in guardrails are not enough

Organizations often assume that native AI guardrails or provider-level controls are sufficient to manage AI risk. But securing AI requires ongoing visibility, oversight, and governance, not just controls configured at deployment. "It’s a misconception that adopting AI is going to solve all your problems,” warns a global financial services CISO.

Security leaders are increasingly recognizing the limitations of these controls as:

• Fragmented and difficult to enforce consistently across multiple AI systems, workflows, and environments
• Ambiguous in terms of accountability due to shared responsibility for AI governance between IT, security, developers, business teams, and third-party providers
• Limited in end-to-end oversight, leaving gaps that stretch from the initial prompt all the way through to the downstream impact of an agent's actions

Securing AI demands more than simple prompt filtering or static policy enforcement. It requires understanding intent, behavior, and context across both human and AI activity.

The next phase of cybersecurity: securing AI

To safely and responsibly adopt AI at scale, organizations need a new operational model for cybersecurity that’s capable of:

• Understanding AI behavior

• Identifying risk in real time

• Maintaining governance without slowing innovation

The CSO of a $10 billion municipal utility organization described the challenge with precision: “We have to move at the speed of innovation and risk, because both are accelerating faster than ever.”

Embrace AI with confidence with Darktrace / SECURE AI

Darktrace has introduced Darktrace / SECURE AI™, a new product within the Darktrace ActiveAI Security Platform™  ,designed to provide enterprise-wide security for AI by applying industry leading behavioral analysis to how prompts, agents, and AI systems are used.

Darktrace / SECURE AI^TM delivers real-time visibility and control across Enterprise and SaaS GenAI prompts, AI agent identities, development and production environments, and Shadow AI - detecting even subtle misuse, misconfiguration, and drift that traditional, rule-based controls simply do not understand. By interpreting context and intent across humans and machines, Darktrace enables organizations to adopt AI at scale without introducing unmanaged risk

What makes this possible is Darktrace’s decade-long maturity and expertise in behavioral understanding and AI-native cybersecurity. Achieved with Self-Learning AI that has been proven across more than 10,000 organizations, Darktrace understands what “normal” looks like for a business, across its users, systems, and now AI, so that meaningful deviations can be detected and acted on before they become incidents.

With one CISO describing Darktrace’s Self-Learning AI as “a leap forward compared to other tools” and another as a “force multiplier,” the technology can interpret ambiguous interactions, understand how access accumulates over time, and recognize when behavior, human or machine, begins to drift.

“Strategically, we’re looking to gain more visibility into how AI is operating across the environment and achieve greater control over what AI should be allowed to access and do,” shared the CISO at a private fund administrator.  

“What I’ve seen from Darktrace / SECURE AI is extremely promising. I have tremendous confidence in Darktrace’s vision for where this is headed and its ability to execute on this new solution.”

‍