For many years, security teams have relied on a common severity scale to decide which vulnerabilities need attention first. One of the people who helped create that scale now believes it has to evolve, because attackers now combine weaknesses in ways a single score can’t capture. That argument runs through episode 4 of the
In the episode, host Shahar Peled interviews Gerhard Eschelbeck, former Chief Information Security Officer (CISO) of Google and
Terra also announced Eschelbeck's appointment to its
A Career Built Around the Vulnerability Problem
Eschelbeck has spent almost 30 years in cybersecurity. He began at McAfee, which he calls a pioneering period, then joined the early team at Qualys before serving as CISO at Google for a number of years. In recent years he has concentrated on the startup ecosystem, working with early-stage founders and the wider security community.
His time at Google shapes how he thinks security ideas mature. He recalls the company working on Zero Trust and passwordless initiatives years before similar concepts reached the open market. He also credits Google’s security organization with an unusual ability to anticipate the next move. That habit of looking ahead runs through his comments on scoring, testing and automation.
How CVSS Started as a Grassroots Fix
As Eschelbeck tells it, CVSS grew out of a practical problem. Vendors and partners were disclosing vulnerabilities without a shared way to express severity. Some labeled issues high, medium or low, while others used severity one, two or three. That left CISOs unable to compare findings or decide which issues needed a patch that day and which could wait a week.
“Well, it started as a grassroots effort, really,” he said. Eschelbeck and two collaborators spent time defining the initial standard, and the result gave the industry what he calls almost a common language for vulnerability severity. Peled called the framework one of the most impressive achievements on Eschelbeck’s résumé, given how long it has been used across the field.
What Does Severity Scoring Miss Today?
Eschelbeck still sees value in CVSS as a shared vocabulary, and he says that need hasn’t gone away. What has changed is what gets analyzed. Security teams no longer assess one vulnerability in isolation. They look at attack graphs, and chaining vulnerabilities together is not necessarily part of CVSS.
“I think what has changed is the chaining of vulnerabilities,” he said. Several individual weaknesses, none of which looks high-profile alone, can combine into a scenario no one had found before. In his framing, the mindset has moved from the severity of one vulnerability to the exploit attack graph. Volume adds to the pressure: “the speed, the scale, the amount of vulnerabilities and risks that we are seeing is unheard of,” he said, putting the change at roughly an order of magnitude compared with a year earlier. AI has also introduced new categories of issues, including prompt injection.
He sees two paths forward: revise CVSS, or build an entirely new system for the AI world. When Peled asked whether he would want to take part in that work, Eschelbeck answered, “Love to be.”
The Visibility Gap Behind Successful Attacks
Peled raised what he considers the central question in security: if adversaries attacked today, what could they achieve? Eschelbeck answered as a CISO would. The biggest challenge, he said, is having a complete picture of an environment that never stops changing.
A few years ago, he studied why companies get hacked. Looking at well-known successful attacks, he found two primary causes. The first was user error, such as missing multi-factor authentication (MFA) that let a stolen password lead to a breach. The second was not knowing what exists in the environment, which lets a stale system become the way in. Asked whether any single dashboard can answer Peled’s question today, he said, “Not really, not really today. We’d love to build it as well, but that’s the next startup opportunity.”
Where Automated Penetration Testing Fits
Peled described traditional penetration testing as project-based, point-in-time and narrowly scoped, usually run once a year or once a quarter against a small slice of the attack surface. Eschelbeck argued that this model no longer fits, because any change to a code base, however small, can affect production right away. To understand attack paths, he said, teams have to examine the whole chain from source code to production.
“AI is the tool to allow us to leverage penetration testing in an automated way,” Eschelbeck said. That testing can run regularly, frequently and automatically, with different degrees of autonomy. He was clear, however, that fully autonomous cybersecurity is not a reality today. “We are better off today with optionality where we take advantage of the AI-based scale together with the human decision-making,” he said. He compared the field’s current stage to where autonomous vehicles were five to ten years ago.
Agents, Mythos and an Expanding Attack Surface
Eschelbeck says advanced models such as
Those same capabilities create new oversight demands. Peled described a
An Open Project for the Industry
For security leaders, the most actionable part of the conversation may be Eschelbeck’s willingness to reopen the scoring question he helped settle. A framework that accounts for chained vulnerabilities and attack graphs would change how teams rank and report risk, and he has said he wants to help shape it. Until that work takes form, his comments point toward continuous testing across the full code-to-production chain and a more complete inventory of what each organization actually runs. The full conversation is available as episode 4 of the AI Native Cybersecurity podcast from Terra Security.
This article was published under HackerNoon's Business Blogging program.
