Equifax automates half of its security tickets with AI

Equifax is using AI to process 19.8 million security alerts and scans per day, with the company saying automation now handles half of its security operations center incident tickets. The effort follows years of remediation after the 2017 breach, whose cleanup costs reached $1.4 billion, and comes as automated attacks shorten the time organizations have to patch exposed systems.

Jeremy Koppen, Equifax’s executive vice president and chief information security officer, joined the company in May 2025 after 13 years at Mandiant, now part of Google Cloud. He says the increase in attacker automation has already changed the volume of external threats.

“With the increase of AI capabilities, we’ve seen an increase in external attacks. I was just looking through the numbers, and we’ve had about a 30% volume spike driven by automation.”

The operational problem extends beyond alert volume. Equifax says the mean time between a vulnerability becoming known and being exploited is shrinking, leaving less room for conventional triage and patching. Its response combines automated analysis with controls that keep humans responsible for consequential decisions.

Recommended reading

Chinese hackers used an IoT botnet to breach NASA and more

Sophia Reynolds • Aug 28, 2026 • 3 min read

AI handles triage, code review and vulnerability fixes

Equifax says AI automatically handles 50% of SOC incident tickets, giving human analysts more time to work on the cases judged most critical. The systems also add context and knowledge to tickets so analysts can begin investigations faster. Koppen stresses that automated remediation does not remove the need for verification.

“We can use AI to help remediate, 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.”

The company’s 2025 security annual report, released in March, describes a certificate-management tool that automatically renews and tests TLS certificates. That directly addresses the expired public-key certificate involved in the 2017 breach, though the report does not establish how much of the overall risk reduction comes from that tool.

Equifax is also adding AI earlier in its software-development process. Security reviews that once took 46 days now take 18 days, the company says. Its AI agents analyze container vulnerabilities and automatically write code fixes, handling more than 213,000 findings annually without slowing delivery. Equifax also says security consultation times are down by 61%.

Those figures are company-reported operational measurements, not independent benchmarks. The supplied reporting does not specify the systems used, the baseline period for the percentages, or the error and approval rates for the automated fixes. That makes the results useful as evidence of deployment scale, but not as a direct measure of security effectiveness.

Equifax is treating AI agents as production identities

The harder problem is controlling agents that can inspect systems, modify code or access networks. An agent capable of finding a vulnerability in Equifax’s environment could also discover weaknesses in its own restrictions, exfiltrate data or use the company’s infrastructure to attack a third party.

Equifax’s answer is to apply older identity and network-security principles more strictly to software agents. The company says agents are limited by identity-based access controls and network boundaries, with access restricted to the resources required for a task. Its security report says manual gatekeeping has been replaced with policy-as-code: every new agent is tested before production, while high-stakes actions require human approval.

“When we’re using agents, we want to make sure we lock down what the agent can get to. That’s key from 20 years ago. I think it’s just more imperative now to have those be implemented.”

Equifax also says it continuously monitors agents for drift or unpredictable behavior. Automated kill switches and instant rollback capabilities are intended to stop an agent before a bad action expands beyond its permitted scope.

“I think about it from a network perspective. 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.”

That model is closer to treating an agent as a privileged production identity than as a chatbot. The practical trade-off is additional policy, testing and approval infrastructure around every automated capability. Equifax is expanding its passwordless strategy beyond its 22,000 employees and contractors to business partners, removing passwords as one route for social engineering while leaving agent permissions and partner access as separate control problems.

New attack path: hidden instructions inside content

Equifax’s attack-simulation team also found that adversaries could embed invisible text prompts designed to trick AI models into delivering malware. The company built a live prevention control that strips those hidden commands before they can cause damage.

That defense targets the content reaching a model rather than the model’s underlying weights. The reporting does not say which model types, file formats or delivery channels are covered, nor does it provide a false-positive rate. Those details matter because stripping invisible text can affect legitimate documents as well as malicious prompt-injection payloads.

Equifax is also using a business-exposure map that combines business data with a quantitative risk engine. It ranks problems according to factors such as whether an asset is externally exposed and how many layers of defense protect it, allowing the company to prioritize patching as the exploitation window contracts.

“We can make sure we’re prioritizing and reducing that risk. 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.”

The approach reflects a lesson that Equifax’s breach made expensive: basic controls still determine whether advanced tooling helps or amplifies damage. Passwordless access, certificate renewal, network segmentation, policy enforcement and human approval are doing as much work here as the AI models. The unresolved question is whether those controls can keep pace when the agents themselves become capable of finding and exploiting the next weakness.