Internet Security Alliance details agentic AI risk in critical infrastructure deployment

“Adoption is outrunning governance in every sector, including the most capable ones. Agentic AI presents a different governance problem than sectors have faced before, and the risk is primarily in the operating envelope around the model, rather than the model itself. Periodic compliance can no longer manage the systems being governed,” ISA president and CEO Larry Clinton said at a Sept.

2 congressional briefing . The briefing focused on the findings of a cross-sector assessment from ISA examining AI deployment across critical infrastructure. Clinton authored blog posts on security gaps and challenges and policy solutions to fill the gaps.

Speakers included Fariba Wells of Kyndryl, Mike Woods of GE Vernova, L3Harris’ Michael Higgins, and Mastercard’s Allisah Abdullah. The speakers provided an overview of AI risk in information technology, energy, the defense industrial base and financial services sectors. Wells said the “core policy challenge” facing AI governance in the IT sector is the introduction of “autonomous digital actors into critical infrastructure.” “While most cybersecurity policies were written for human users in traditional software, AI is quicker than what we're seeing than cybersecurity can keep up,” Wells said.

She pointed to a July 21 security incident involving an OpenAI model gaining unauthorized access to the Hugging Face platform during a performance evaluation as an example. “Agentic AI is a new class of operational actor that does not fit neatly within the workforce management or traditional IT service management, and agents can make those decisions and take actions, and they can scale,” Wells said. She said there is “no end-to-end governance for agents” in the IT sector, arguing the need for human-in-the- loop requirements and risks from shadow AI and missing lifecycle controls.

“As Congress is exploring this, we need to make sure that we take a risk-based interior governance transparency approach using a common AI security baseline,” Wells argued. “AI is advancing faster than existing cybersecurity frameworks, and we need to make sure that the industry and government are all working together so we can build trust and ensure innovation doesn't outpace security,” she said. Woods said AI technologies are being deployed in “every layer” of the energy sector, including product development, manufacturing, supply chain, operations and field services.

He argued that there is a “huge trust gap” in how the energy sector is deploying AI technologies, highlighting the need to adopt secure by design principles and voluntary cybersecurity testing as options to increase trust. “The outcome that we want to get to with this, at least as the energy sector, is we must make the AI deployments across the sector trustworthy enough for all of the systems that America depends on for energy,” Woods said. “The real takeaway is AI security in the energy sector is a reliability, resiliency and a national security issue,” Woods said.

On the defense industrial base, Higgins said, “As cybersecurity defenders, we are quickly realizing that we only have a short period of time to address systemic security issues before AI-enabled attacks begin to scale, and these attackers will quickly gain capability as AI models improve.” He said policymakers need to take a “collective response” to AI-risk mitigation, with a focus on providing access to frontier AI models with advanced cybersecurity capabilities. Mastercard’s Abdullah also argued the need for AI governance in the financial sector, saying, “AI is not going to slow down. Adoption is not going to slow down.

What responsible organizations have to do is know where AI is operating.” She said, “We have to give AI limited authority. We have to put controls outside of the model. We have to make sure we preserve human accountability.” On the sidelines of the briefing, Clinton told Inside Cybersecurity , “The biggest takeaway is that we need to do things very differently than we have in the past, and we need to start doing it now.