In my previous column, I offered some suggestions to help security teams avoid being blindsided when AI applications are moved into production. In this piece, I’d like to offer some thoughts on what is required for security teams to efficiently and effectively incorporate AI applications into the operational security workflow. While there has been much hype around AI applications, many security teams struggle with securing, monitoring, and defending them for a variety of reasons.
While not an exhaustive list, I’ve put together 12 practices that I’ve found helpful for incorporating AI applications into the operational security workflow:
- Visibility: We cannot defend what we cannot see. As such, visibility is really the most fundamental of building blocks when it comes to securing AI applications. Beyond just awareness of and inventorying AI applications, visibility can help us identify exposures of sensitive data, vulnerabilities, deficiencies in controls, fraud, abuse, attacks, and other issues. This makes continuous visibility an extremely important ingredient when it comes to incorporating AI applications into the operational security workflow.
- Understand Risk: If we’ve taken visibility seriously, we’ll have good data around risk. That data can be used to scientifically understand risk, rather than playing a guessing game. Beyond just a snapshot in time, understanding risk can be done on an ongoing basis in near real-time to allow the security team to more precisely evaluate the risk that one or more applications present to the enterprise. This makes understanding risk another helpful tool when it comes to incorporating AI applications.
- Build Trust: Discovery is an important part of the visibility piece discussed above. The data generated during discovery can be used to catalyze building relationships between the security team and other important stakeholders, such as application owners, product management, developers, and others. In time, these relationships can mature and a fair amount of trust can be built. This trust will serve the security team very well.
- Leverage Trust: The trust built in the previous step often facilitates the security team involving itself much earlier in the software development life cycle (SDLC). That makes it much easier to incorporate AI applications into the operational security workflow, which is good news for the overall security posture of the enterprise.
- Telemetry: If discovery and visibility have been done right, there should be a fair amount of telemetry being generated. It is important that this telemetry data thoroughly cover the AI applications and the infrastructures they’re built upon. This involves generating telemetry from inspecting the AI layer, API layer, and application layer and ensuring that data flow to the SIEM, SOAR, or preferred system of record. Having eyes deep into the application and its infrastructure and making that telemetry available to the security team for analysis, investigation, response, and other steps is extremely important when it comes to properly securing AI applications.
- Process: While it may be the least sexy one of these points, it is important to develop processes and procedures around securing AI applications, just like we do for all other areas of security. This provides important guidance to the security team and allows them to act and react more agilely – something very important when trying to handle the latest hot potato thrust upon the security team.
- Enforce: The best controls in the world won’t do an enterprise any good if they can’t enforce those controls. Thus, ensuring that the security team has the ability to easily implement and enforce controls across the various different environments where applications run is an extremely important component when securing AI applications.
- Preventive Controls: Good preventive controls across the enterprise are going to help security teams with anything, including incorporating AI applications. It is important to remember that these preventive controls should include protection against abuse, fraud, DDoS, malicious automated attacks, and other threats. As such, security teams should ensure they have adequate preventive controls in place, along with the ability to augment and improve those controls in an agile manner.
- Detective Controls: Continuous Security Monitoring is an important function within any operational security workflow. These detective controls serve as an important partner to preventive controls. However, they require that the requisite visibility is in place and the appropriate telemetry is flowing to the system of record. Assuming this has all been set up properly, it greatly facilitates the security team in its efforts to secure AI applications.
- Investigation: In the event that a security issue is uncovered, the security team will need to be able to analyze and investigate the corresponding data (logs, events, alerts, etc.). Ensuring that this level of investigation is possible will be another important part of incorporating AI applications into the operational security workflow. It isn’t enough to collect the requisite telemetry – that data need to be available for interrogation by the security team.
- Mitigation: In the event of a security issue, once the investigation is complete (or at least far enough along), the security team will need to be able to respond, remediate, and recover. This requires that the reach back into the application, API, and AI infrastructure be in place for this to happen. This is something that enterprises often forget about, until they learn the hard way that they can’t mitigate a security issue once they become aware of it.
- Iterate: Although most of the previous points are more fun than deriving lessons learned, documenting findings, and implementing those lessons learned on a continual basis, these are extremely important. Only through continuous improvement can a security team keep pace with the rapidly changing threat landscape. Enterprises that learn how to iterate successfully will be more nimble when it comes to making the adjustments required to properly secure AI applications as time goes on.
AI applications moving into production does complicate things for security teams. That being said, there are steps security teams can take to ease the burden. Hopefully, security teams will be involved earlier in the software development life cycle (SDLC) in the future. Until that time, however, they will need to take steps to proactively prepare to incorporate AI applications into the operational security workflow.
Learn More at the AI Risk Summit | Ritz-Carlton, Half Moon Bay
