Crest CEO: 'Basics shouldn't be abandoned in the race to deploy AI'

As an international non-profit representing the global cybersecurity industry, Crest has been closely tracking the rapid evolution of AI -- balancing its tremendous potential as a defensive tool against the mounting threats posed by malicious actors.

Now in its 20th year of operation, Crest has taken significant strides to address this transitional period head-on, transitioning the industry from voluntary AI commitments to independently verified standards.

This momentum kicked off in June with the launch of the organization's AI Charter, which received signatures from a global cohort of over 100 cybersecurity companies committed to responsible AI use. Building on that foundation, July saw the milestone launch of Crest's AI-Enabled Penetration Testing accreditation, which establishes independently assessable requirements for secure and appropriate AI integration. In August, Crest followed up with its Security Testing of AI standard to provide rigorous evaluation criteria for LLM-enabled systems, closely mirroring warnings from industry leaders and open letters calling for collective action.

Against this backdrop, we caught up with Crest CEO Nick Benson to discuss the driving forces behind these new standards, how service providers can maintain critical human oversight and what the future holds for global regulatory alignment.

This interview has been lightly edited for clarity and length.

TT: Your research shows that 76% of cybersecurity providers increased their use of AI in the past year, yet you note that "recognized standards for demonstrating responsible AI use have not kept pace." What are the most dangerous gaps you've observed?

NB: The strongest demand signal we are seeing for such standards is from the client side, where sometimes conflicting pressure is being put on service providers to demonstrate they are adopting AI rapidly to drive down costs, but also being highly cautious about that adoption.

Historically, though, service providers have had no clear standard against which to demonstrate their capability and give clients and prospective clients confidence. Providers are exploring increasingly sophisticated uses of AI, including greater automation within security testing, but the maturity of the controls surrounding that use can vary.

Without clear policies and safeguards, this can introduce risks in areas such as the handling of sensitive client data, the use of public AI tools and the appropriate oversight of AI-generated outputs. 

In addition, some providers may otherwise rely on unvalidated automated outputs. These can generate false positives or miss logical flaws. This shows why human oversight is still so important.

TechTarget: Your new AI-Enabled Penetration Testing standards provide "practical, independently assessable requirements" for responsible AI use. Can you walk us through what specifically gets tested during accreditation?

Nick Benson: When we accredit providers, we are specifically evaluating transparency, operational controls and safety boundaries. For example, are they clear with their clients about what AI capabilities they will use, how they will use them and how the client's data will be protected? We also look into the use of human oversight in governing how AI is used in service delivery.

An applicant must undergo organizational due diligence checks and pass the requirements of our general penetration testing accreditation and the AI-specific supplement in order to be accredited. This includes evidence of policies, methodologies, processes, experience and skilled people relating to the service. Failing to provide sufficient evidence of capability at any of these stages would prevent an applicant from being approved.

Unlike voluntary agreements, this formal accreditation integrates directly into our existing complaints and discipline processes. This provides an incentive for service providers to maintain the high standards expected of them and a route for customers to raise complaints where appropriate, which can be subject to enforcement action.

TT: The same AI capabilities that help defenders identify vulnerabilities also help attackers exploit them faster. What is the consensus on whether AI is currently tilting the advantage toward attackers or defenders?

NB: Our industry has always involved a cat-and-mouse game between adversaries and defenders. Arguably, the biggest advantage an adversary has is the lack of legal, regulatory or ethical constraint. This means that they can go full throttle in experimenting with its capabilities at a pace that can exceed those that are more appropriately bound.

In answer to your question, though, we are seeing rapid development and deployment of AI in a defensive context, as it is clear that to fight AI, you need to use AI. The consensus across our working group is that both enterprise and cybersecurity providers are having to deploy AI at pace, working in parallel to bring in the controls that help mitigate the worst of the unintended consequences.

It's really important to remember, though, that the basics of cybersecurity mustn't be abandoned in the race to deploy AI solutions. In many cases, we are still seeing the greatest harm being from AI-enabled adversaries more rapidly exploiting basic cybersecurity vulnerabilities rather than novel or zero-day exploits.

So, yes, we are seeing an inevitable rush to deploy AI on the defensive side to counter AI-enabled adversaries, but not at any cost. Control and technical innovation can operate side by side. For this to be the case, governance and verification need to mature at the same pace as AI tools' capabilities.

TT: You previously mentioned that buyers, regulators and procurement teams are demanding greater accountability. What specific regulatory developments are driving this demand for AI accountability? Looking ahead 12 to 24 months, do you anticipate independent AI security certifications becoming a mandatory requirement rather than a competitive differentiator?

NB: As I've mentioned, regulatory adjustments typically lag technological innovation, but due to the geopolitical focus on AI, the gap is actually closing surprisingly fast. Most governments now appear to be looking at how to regulate AI development, release and security. In most cases, that is still in the form of announcements, guidance and advice. The industry, in parallel, is driving self-regulation to demonstrate capability to their clients, which is where CREST has seen the greatest action to date.

The critical infrastructure sectors tend to be first to research and regulate emerging risks, given the higher inherent risk and often better-funded regulators. We are seeing this in financial services in particular, but also, for example, region-wide through the EU Cyber Resilience Act. Meanwhile, there are national AI directives like the US NIST AI Risk Management Framework and international standards like ISO/IEC 42001, which specify requirements for Artificial Intelligence Management Systems.

Enterprise procurement teams now routinely ask how service providers govern AI and protect client data. Over the next one to two years, independent verification will not be more than a competitive differentiator. I believe it will actually be a standard procurement requirement.

TT: How do you establish a globally applicable standard across regions with very different AI regulations, and how do we avoid fragmented security standards that create compliance complexity?

NB: Core to Crest's mission is the harmonization of standards internationally, helping manage cross-border risk in a sector where national boundaries have little effect. In other words, the adversaries operate internationally and therefore will exploit the gaps between national regulations and law enforcement if international efforts to coordinate fail.

Our model is to work closely with national regulators, accepting that sovereign decisions will always be made, but that if we can be the glue between them, we can drive a more harmonized approach. This is what we are doing, for example, with the International Coalition on Cyber Security Workforces (ICCSW), where we are supporting the common principles and frameworks behind professionalizing the cyber workforce.

We would argue that there are many common characteristics of AI security regulation, but that the optimal model is for there to be a common baseline. This is why we set a globally applicable standard focused on fundamental principles and controls, aligned with other international frameworks such as ISO/IEC 42001.

We know and accept that national authorities need to tailor the standards to some degree locally, but relying on a common baseline can expedite standard setting, reduce cost and ultimately be a more effective weapon against adversaries.

James Walker is lead editor at TechTarget Cybersecurity.