About three years ago at the Hartford, Conn.-headquartered Aetna, a health plan with more than 37 million consumers, organizational leaders set out to create new security measures for its mobile and web applications that would aim to transform existing controls.
At the core of the initiative was being able to monitor user behavior in real time, says Jim Routh, the chief security officer (CSO) and global information security function leader for Aetna. “As it turns out, security is evolving pretty quickly into a model-driven security realm,” he says in a recent interview. Routh explains that model-driven security centers around frontline security controls in which algorithmic models determine things such as: how much access to give to a consumer, an employee, or a privileged user; whether something running on an endpoint device is malware; or whether a phishing email is being sent through the email infrastructure.
In many cases, notes Routh, it’s the models that are driving security controls; at Aetna, there are 200 models in production today that are doing just that. “And we do a lot of manipulation of the models, which is evolving cybersecurity and physical security practices from conventional to unconventional controls,” he says.
In the interview, Routh spoke to Healthcare Informatics about how Aetna has been able to put these controls in place, why behavioral-based security is so important and more. Below are excerpts of that discussion.
Tell me about your plan at Aetna to monitor users’ behaviors in real time. How did it all begin?
Three years ago we hired a chief data scientist to be dedicated to security, and this was someone with nine years of experience at the NSA (National Security Agency). We asked him to build a large data lake for security, enterprise-wide and to scale, that would be something we could use off-the-shelf to do better cyber hunting, and figure out patterns and anomalistic behaviors across a wide swath of data, and develop broad models for some of our payment business segments.
And he did that exceptionally well. In two-and-a-half years’ time, he had 110 models running in production and it was exactly what we asked him to do. The irony is that during that time we deployed eight other implementations of technology platforms that had embedded unsupervised machine learning to drive controls.
Can you give some examples of these models?
The first and most significant example from a consumer perspective is when we moved into what we call continuous behavioral-based authentication. And there are two parts to that, with one being that it is continuous. Every time we have designed authentication for any kind of technology, it was an event at the front end of an electronic interaction. So in a web app, you provide the user ID and password, you are in and trusted, with full access to the system. And there is little monitoring thereafter since you’re a trusted entity. Remember, binary controls for authentication can be defeated since they are based on assumption that only the end-user has the information. And that used to be the case with passwords, but it isn’t anymore.
In 2016, three billion credentials were harvested just based on public breach data. Shape Security [a Mountain View, Calif.-based security company] did the analysis based on public breach data, and they believe that the number is actually closer to 10 billion credentials. And there are only 320 million people in the U.S., so the assumption that you are the only one that has your password is no longer valid.
There is actually a tool called Sentry MBA that bad actors could use. So let’s say that they harvest 10,000 e-mail credentials from Yahoo, for example, take those credentials (user ID and password combination) and try it out on any other domain that they wish—and they can do it through a script in the Sentry MBA tool—they will get a 2 percent hit, so they will own 200 accounts from the 10,000 credentials that they attempted. And the reason that 2 percent hit is because you and I, like everyone else, can’t remember passwords for 100 different sites and mobile apps. So we use the same password. So 2 percent of the time you will get a hit just by using the same user ID and password from one domain to the other.
We are reaching the point of when the credential availability will be unlimited, and when that happens, there is no friction stopping a criminal to do this at scale. A growing percentage of authenticated log-ins in an enterprise are done by somebody using someone else’s credentials. So the obsolesces of passwords will continue to grow and grow significantly. It will take a decade to swap out login credentials across all of the enterprises, since that doesn’t happen overnight; 99 percent of authentication today is done through passwords. But in order to solve this problem, we have to recognize that authentication moves from an event to the front end of an interaction to a continuous process. And we are using behavioral attributes gathered electronically, and we apply a risk score to that, and that risk score notifies the application how much access to provide throughout the interaction.