Recent cybersecurity incidents have highlighted how weaknesses in software and third-party dependencies can quickly become customer-facing risks and the security challenges facing enterprises that depend on complex software ecosystems to deliver customer experiences. A vulnerability several layers below an app or digital service can disrupt transactions and expose customer information, and service disruption can provide attackers with an opportunity to target employees and customers, damaging confidence in a brand.
For enterprises, responding to these threats requires more than identifying vulnerabilities and applying patches. It means understanding the software running across their application portfolios, improving the speed of security testing and finding ways to remediate problems without disrupting customer-facing services.
James Labocki, Senior Director of Product Management at Red Hat, told CX Today that AI harnesses could play an increasingly important role in that effort, helping security teams discover vulnerabilities and attack paths by giving models the context and tools needed to test applications more effectively.
The accelerating pace of vulnerability discovery and exploitation demands a fundamental rethink of how organizations build and maintain software.
"It's vital" that organizations can transparently communicate to customers what their applications are running on and demonstrate that those applications are secure,” Labocki said.
"Being able to assure them that the software that's running your business is actually safe and has proper integrity and can be trusted is extremely important."
Why AI Security Harnesses Need the Right Context
The arrival of increasingly capable AI models is changing the economics of vulnerability discovery. Labocki pointed to the “Mythos moment” in April, when Anthropic's announcement around its frontier model brought renewed attention to the potential for AI to identify software weaknesses and develop exploits.
"For a long time, organizations sometimes bolted on security as opposed to building it in," making security "somewhat of an afterthought" and slowing down the ability to identify and address problems. But as "time to exploit" compresses, enterprises must reconsider how security is embedded into their development and operational processes.
One emerging response is the use of AI harnesses for adversarial testing of application codebases.
An AI harness provides the framework around a model, connecting it to relevant tools, instructions, application data and security controls. In a software security context, this can include supplying an application's architecture, defining its threat model and directing the model towards particular components or potential attack paths.
The model can then use that context to investigate weaknesses and help security teams identify vulnerabilities that conventional testing might miss.
"It's not necessarily the model that's the most important. A lot of it is the context and the harness that's used," Labocki said.
"Bringing the appropriate threat model about an application in that context can really help the model discover new vulnerabilities and attack vectors more effectively."
The approach is already gaining traction among enterprises and security providers, with organizations building their own harnesses to test their applications.
