AI agents spark new identity security fears for 2026

Identity security specialist One Identity has set out five predictions for 2026, warning of rising AI-driven breaches, proof-based supply-chain oversight, and stricter control of non-human accounts and digital identities.

The company said identity will sit at the centre of a new phase of cybersecurity and digital governance. It highlighted growing interdependence between human users, software bots and AI agents across large organisations.

Its forecast points to intensifying supply-chain attacks, new forms of AI abuse, and regulatory pressure in Europe around digital identity wallets.

AI agents at risk

One Identity expects the first major security breach that starts with an AI interface and results in privilege escalation. It said many organisations are granting operational powers to AI systems that can already execute commands, orchestrate workflows and change production settings without human involvement.

Alan Radford, Global Strategist at One Identity, said the shift from passive AI assistants to active AI agents is outpacing many security programmes.

"The AI goldrush has begun. AI assistants are graduating to AI agents that go beyond informing to also take action, becoming operational gatekeepers that are authorized to execute commands, orchestrate workflows and even change production configurations without human interaction", said Radford.

He continued, "As the rush to invest in AI continues, many will run before they can walk, outpacing the capabilities of their current security programs. Undeniably, we are on the verge of witnessing the first high-profile breach to emerge from this new attack surface, where over-privileged AI will be exploited.

"Prompt-injection and response-time attacks will target the workflow between user and agent, triggering unauthorized workflows, revealing and manipulating sensitive data, unintentional changes in configurations and over escalation of privileged permissions and roles. When this does occur, it will expose how fragile current privileged boundaries have become where AI is authorized to act autonomously on human requests."

Radford said this shift will change defensive priorities inside security teams. He expects a focus on the identities that reside within AI systems, including machine and agent identities.

He said, "The direct consequence is a shift in defensive priorities. Business will have to move from securing identities that use AI to securing identities within AI itself. The bigger picture goes beyond securing human identities that use AI, to securing the machine and agent identities that live inside AI systems themselves.

"Guardrails will evolve to log not just what the AI did, but who or what process triggered it with a fully traceable chain of custody. Organisations need to demonstrate provable accountability of agentic actions, binding every automated decision to distinct human ownership. In other words, "AI governance" is no longer an abstract compliance term, it has become an essential frontline security control."

Proof in supply chains

The company also predicts more pressure on supply-chain security. It expects third-party compromises to grow as attackers seek weaknesses in the trust relationships between organisations and their vendors.

Regulators are pushing for continuous verification of security controls. One Identity said boards will demand real-time access evidence rather than annual attestations.

"The next evolution of governance will be proof-based," said Stuart Sharp, VP of Product Strategy at One Identity. "Boards and regulators won't accept annual checks anymore; they'll want real-time evidence of who granted what access, when, and why. Identity will become the shared control plane between organizations and their suppliers. 'Trust but verify' will become 'prove and continuously enforce.'"

Non-human accounts

One Identity sees a mounting "non-human identity" problem. It said bots, service accounts and digital agents outnumber employees on many enterprise networks, often by a factor of 50 to one.

Many of these accounts continue to exist long after their business purpose ends. The firm expects identity teams to introduce stricter lifecycle rules, including kill switches and automatic expiry policies.

"We've reached the point where the biggest insider risk doesn't have an employee ID," said Robert Kraczek, Global Strategist at One Identity. "Non-human identities need the same governance discipline as human users, including ownership chains, expiry dates, and emergency shut-off mechanisms."

Model poisoning fears

The company warns that AI model poisoning will gain prominence as organisations train and fine-tune models on internal data. Attackers may aim to tilt model behaviour rather than shut systems down.

It said subtle manipulation of training inputs could change analytics, automate flawed decisions or bias outcomes in ways that appear legitimate.

"AI assurance will become inseparable from identity assurance," said Nicolas Fort, director of product management at One Identity. "Organisations will need to track not just who accessed a model, but who influenced it and who still has access. Every training event, prompt, and parameter change must be tied to an authenticated identity."

Wallets and eIDAS

The spread of EU digital identity wallets under eIDAS 2.0 is another area of focus. One Identity expects these wallets to normalise verified external credentials in many sectors.

The company said enterprises will need systems that can accept government-issued digital IDs while enforcing internal security policies that limit access.

"For years we've talked about federated identity, and now it's becoming citizen-driven," added Stuart Sharp. "In 2026, users in the EU will expect to 'bring their own' digital ID wherever they go., and enterprises will need to be able to accept those credentials without surrendering their own governance controls."

Beyond the five headline trends, One Identity expects broader changes in identity practice. Its outlook covers AI-based "immune systems" for identity, a renewed focus on data access governance, a shift from cyber recovery towards wider cyber resilience, and a return to identity and access management fundamentals.