Non-human identities are quietly proliferating across enterprises, expanding the attack surface at an exponential rate. Threat actors are paying attention, experts warn, and CISOs should do the same.
For example, in March 2025, a third-party GitHub Action used in more than 23,000 repositories began dumping users’ secrets into workflow logs in public repositories. No one was phished, and the attacker never logged in. Rather, the tj-actions compromise hinged on a personal access token (PAT) that a threat actor stole from a bot.
“So how did the attacker get that token? Well, that was from another breach of another GitHub Action called reviewdog,” said Dhivya Balasubramanian, cybersecurity manager at Southwest Airlines, speaking at ISC2 Security Congress 2026, in Aurora, Colo. The attacker had, in turn, compromised reviewdog using an automation token they stole from SpotBugs, an open-source static analysis project.
“You see the ripple effect here, right? It's a chain of trust, and every single identity in that chain could be a potential entry point.”
That chain of NIHs grows longer by the day, introducing growing risk to enterprises. By mid-2025, machine identities outnumbered human identities by as many as 144:1, a 44% year-over-year increase, according to research from Entro Labs. Even organizations that rigorously govern human identities — promptly adjusting access controls when employees leave or change roles, for example — often overlook service accounts, API keys, PATs, CI/CD runners, workload identities and AI agents. And unlike human identities, many NHIs have indefinite access, static secrets and over-scoped permissions, with little to no human oversight or ownership.
“Hugging Face is a perfect example,” Balasubramanian said, referring to an incident in which OpenAI agents escaped their test environment and breached another company’s infrastructure. “A good identity security system at Hugging Face could have prevented the level of access that OpenAI's agent had to its systems — even if it could not have prevented the initial breach.”
Break the NHI attack chain
Balasubramanian and co-presenter Vikas Gattu, tech lead at Southwest Airlines, described the following six-stage NHI pipeline attack chain.
- Discover. Find NHIs, repos, pipelines and cloud assets.
- Expose. Find or instigate leaks in code, logs or builds.
- Acquire. Steal a valid token, secret or key.
- Exploit. Use the NHI.
- Escalate. Acquire additional privileges, identities or downstream reach.
- Persist. Maintain access through backdoors or persistent keys.
“The key principle of governance is to break the chain as early as possible,” Gattu said, adding that early intervention means less data exposure and greater control.
He urged security teams to immediately start inventorying every NHI in their environments, including CI/CD runners and AI agents, beginning with their top three pipelines. Document each NHI’s purpose and assign it a human owner.
Next, Gattu advised, turn on secret scanning and log masking, find tokens older than 90 days, and replace long-lived tokens with short-lived credentials.
“Say you lock a door to your house and hide the keys under the doormat and hope that the bad actor won’t find it — that's similar to using static keys,” Gattu said. “We have to use modern methods like [OpenID Connect] policies and workload identity federation.”
As a longer-term goal, Balasubramanian and Gattu recommended building at least one detection signal per kill-chain stage, such as honeytokens, secret scanning alerts, unexpected token minting, unusual API calls and pipeline configuration edits.
“The attacker has to successfully navigate through all six stages to successfully exploit us,” Gattu said. “But for us, we only have to detect them at one stage and stop them.”
“Do not let your AI agents impersonate human or non-human identities”
When it comes to NHI security, AI agents raise the stakes, Balasubramanian warned.
“If there is one thing to take away from here, it’s this: Please do not let your AI agents impersonate human or non-human identities,” she said. In other words, human users shouldn’t share their own credentials with AI agents to let them act autonomously on their behalf. For one thing, most human identities have excessive privileges, which AI agents will intentionally or unintentionally exploit, Balasubramanian said.
Secondly, good governance requires that logs accurately reflect who or what performed a given action, and an AI agent impersonating a human identity muddies those waters. Agents need purpose-built identities and properly scoped tokens, Balasubramanian said, adding that identity security is the only way to control their access and prevent related breaches.
“If there's one thing I would recommend — if identity security was not your priority, was not even remotely in your budget — now would be a good time to start putting identity front and center,” Balasubramanian said.
Alissa Irei is an Informa TechTarget senior reporter covering cybersecurity news.
