The credit bureau’s CISO explains how AI is helping the company defend against attacks, secure code, and accelerate fixes — and why security fundamentals and controls are more vital than ever.
For nearly a decade, Equifax has been dealing with the aftermath of one of the worst cybersecurity breaches in US history, racking up $1.4 billion on cleanup costs.
Among the mistakes that led to the breach were a mismanaged patching process, an expired public-key certificate, and poor governance. One company Equifax brought in to help was Mandiant, now part of Google Cloud.
Today, Equifax, like all enterprises, is facing a different monumental cybersecurity challenge, one that sees AI facilitating attacks at a scale never seen before, altering companies’ security postures with its rapid internal adoption, and, now, with the latest advanced models, showing its potential to escape controls and attack other companies.
The person leading Equifax’s response to this challenge is EVP and CISO Jeremy Koppen, who spent 13 years at Mandiant.
Koppen came on board in May 2025, just in time for the AI threat to start getting real.
“With the increase of AI capabilities, we’ve seen an increase in external attacks,” he tells CSO. “I was just looking through the numbers, and we’ve had about a 30% volume spike driven by automation.”
Meanwhile, if there’s a known vulnerability out there, companies have less time than ever to patch it, Koppen notes. “The mean time for an exploitation of a vulnerability is shrinking with the rise of new technology,” he says. “So that’s top of mind.”
To deal with the growth of AI-related risks and threats, as well as all the traditional threats still out there, Equifax is doubling down on basic cybersecurity hygiene and using AI itself.
For example, Equifax is expanding has its passwordless strategy, which covers all 22,000 employees and contractors, to its business partners as well. “That’s been a critical piece, where we know we don’t have that social engineering aspect,” Koppen says. “Definitely having passwordless has been a big implementation for our team.”
Also this year, Equifax has rolled out a map of business exposure, which uses a quantitative risk engine to examine risk data and business data to figure out where the problems are and how they could impact the business.
That includes addressing the shrinking patching window, he says.
“We can make sure we’re prioritizing and reducing that risk,” he says. “Maybe an asset is externally facing and that would increase the risk. But if it’s shielded by multiple layers of defense in depth and controls, then you get additional levers in what you can do.”
Putting AI to work
To help its security teams contend with the rising number of alerts and scans, which now number 19.8 million per day, Koppen says, Equifax is turning to AI.
Today, 50% of security operations center (SOC) incident tickets are handled automatically, enabling human analysts to focus on the most critical ones. AI also provides knowledge and context to help analysts hit the ground running and reduce the time it takes to process them.
“We can use AI to help remediate,” Koppen says, “but it’s not replacing that human in the loop. You still need that verification to ensure that you know this is correct. And if we fix this piece, what are the additional impacts it may have.”
Automation is also used to help other areas of security. For example, that certificate problem that contributed to the 2017 breach? According to the company’s annual security report, released in March, Equifax has launched a certificate management tool to automatically renew and test TLS certificates.
Koppen is also responsible for the security of the software Equifax produces, an area that’s also being transformed by AI.
“In a standard process, you’d have a design diagram of how things would work and there’s integration with security, and you go back and forth,” he says. But with AI accelerating the pace at which attackers can probe code for vulnerabilities, there’s less time to address code security issues, he adds.
“We can use AI in that process as well,” Koppen says, noting how AI has helped Equifax perform code review earlier in the design process, ensuring adequate guardrails are built in and reducing the time that process takes.
“It used to be 46 days and now we’re down to 18,” Koppen says. “It’s great to be able to save that time but have a result that we’ve verified with a human in the loop and with the security guardrails.”
Overall, according to the company’s annual security report, security consult times are down by 61%, and Equifax’s AI agents analyze container vulnerabilities and automatically write the code fixes. The system is now handling more than 213,000 findings annually without slowing down delivery.
The report also mentions a new attack vector related to AI. The company’s attack simulation team discovered that adversaries could embed invisible text prompts to trick AI models into delivering malware. Equifax built a live prevention control to strip out these hidden commands before they can do any damage.
Watching out for rogue AI
Then there’s the flip side of AI. Having an AI agent that can discover new security vulnerabilities is all well and good when it’s working as expected.
“We want to be able to do that review on our own environments, to make sure we’re identifying any potential vulnerabilities before they’re announced to the public,” Koppen says. “Because we know attackers are going to be doing that.”
Unfortunately, as frontier labs have found out, an agent that’s good enough to find a weakness in your environment might also be able to find a weakness in its own guardrails and controls.
But there are techniques that companies can use to protect themselves, Koppen says.
“When we’re using agents, we want to make sure we lock down what the agent can get to,” he explains, referring to identity-based control, which is not new. “That’s key from 20 years ago. I think it’s just more imperative now to have those be implemented.”
And companies have long had to have protections in place to make sure that data wasn’t being exfiltrated — or that the company’s infrastructure wasn’t being used as a staging ground for attacks against third parties.
Today, of course, there’s a new AI twist to it.
“I think about it from a network perspective,” Koppen says. “That’s key, that’s the first control we have, making sure that we’re locking down the agent so it can’t escape outside of that zone.”
According to the company’s security report, manual gatekeeping has been replaced with policy-as-code, with every new agent automatically tested before production, with human-in-the-loop approvals for high-stakes use cases. In addition, there’s continuous monitoring to stop a model if it begins to drift or behave unpredictably, and automated kill switches with instant rollback capabilities.
Koppen says he’s been watching the situation with the escaping agents from OpenAI and Anthropic. Any sources of information about how to protect the environment is good, he says.
“You know, that’s been an exciting thing, with the community very open and communicating and sharing,” he says. “It’s been a really refreshing thing, just seeing that collaboration.”