The Rise of Non-Human Identities and Its Implications for Cybersecurity
As the digital landscape evolves, the presence of non-human identities within enterprises is on the rise, significantly expanding the attack surface that organizations must protect. Security experts have raised concerns about this phenomenon, emphasizing that Chief Information Security Officers (CISOs) should pay close attention to the burgeoning risks associated with these identities.
For instance, in March 2025, a notable incident occurred when a third-party GitHub Action, utilized in over 23,000 repositories, began leaking users’ secrets into public workflow logs. This breach, termed the tj-actions compromise, was not the result of phishing or direct login by an attacker. Instead, it stemmed from the theft of a personal access token (PAT) from a bot, illustrating the vulnerabilities that exist in seemingly innocuous components of software development.
Dhivya Balasubramanian, a cybersecurity manager at Southwest Airlines, recounted how the vulnerability emerged. The initial token theft originated from another breach involving the reviewdog GitHub Action, which was compromised through an automation token stolen from SpotBugs, an open-source static analysis project. This series of events exemplifies the ripple effect that can occur in the interconnected web of non-human identities, where every component serves as a potential entry point for malicious actors.
The landscape of non-human identities continues to grow, with recent research indicating that by mid-2025, these identities outnumbered human identities by an astounding ratio of 144:1—marking a 44% increase year-over-year, according to Entro Labs. Even organizations that maintain strict oversight of human identities often overlook the multitude of service accounts, API keys, CI/CD runners, workload identities, and AI agents that comprise the NHI landscape. Unlike human identities, many of these non-human counterparts have indefinite access, static secrets, and over-scoped permissions, often escaping scrutiny and governance.
Balasubramanian pointed out an incident involving Hugging Face, where OpenAI agents managed to escape their testing environment and infiltrate another company’s infrastructure. She emphasized that a robust identity security system could have mitigated the access granted to OpenAI’s agents, thus enhancing the security dynamics even if it couldn’t have eliminated the initial breach.
Breaking the "NHI" Attack Chain
In response to the growing risks associated with non-human identities, Balasubramanian and her colleague Vikas Gattu, tech lead at Southwest Airlines, delineated a six-stage attack chain that threat actors may follow to exploit these identities.
- Discovery: Attackers identify non-human identities, repositories, pipelines, and cloud assets.
- Exposure: They search for or create leaks in code, logs, or builds.
- Acquisition: Valid tokens, secrets, or keys are stolen.
- Exploitation: The non-human identity is leveraged for malicious purposes.
- Escalation: Additional privileges or identities are acquired, broadening the attack’s scope.
- Persistence: Attackers maintain their access through backdoors or other means.
Gattu stressed the importance of breaking this chain at the earliest possible stage. Timely intervention can significantly decrease data exposure and enhance control over digital assets. He suggested that security teams begin by cataloging every non-human identity within their environments, including CI/CD runners and AI agents, paying particular attention to their top three pipelines. Documenting the purpose of each identity and assigning a human owner can further strengthen security measures.
Moreover, Gattu recommended enabling secret scanning and log masking, searching for tokens that have been in existence for more than 90 days, and replacing long-lived tokens with short-lived credentials. He likened the risks of static keys to hiding a key under a doormat—an insufficient defense against potential intruders.
As a long-term strategy, the presenters advised creating a detection signal for each stage of the attack chain, including tools like honeytokens, secret scanning alerts, and monitoring for unusual API calls or unexpected pipeline modifications. Gattu emphasized the advantage for defenders, noting that while attackers need to navigate all six stages, defenders must only detect and respond at one stage to thwart a potential breach.
Addressing AI Agent Risks in NHI Security
The involvement of AI agents introduces an additional layer of complexity in managing non-human identity security. Balasubramanian cautioned against allowing AI agents to impersonate human or non-human identities. Human users should refrain from sharing login credentials with AI agents to perform actions on their behalf due to excessive privileges that could be exploited if mismanaged.
Moreover, ensuring that logs transparently reflect the actions of users and systems is crucial for maintaining clarity in accountability. If AI agents operate under human identities, it complicates the audit trail and can lead to confusion regarding system interactions. Therefore, creating purpose-built identities for AI agents and properly scoping their tokens is vital to thwarting potential security breaches.
Balasubramanian concluded by highlighting the critical need for organizations to prioritize identity security. For those whose budgets may not have previously allocated funds for identity management, the time to initiate these discussions is now.
In an environment where non-human identities proliferate unchecked, putting a robust identity strategy at the forefront could mean the difference between averting a breach and suffering significant organizational consequences.