Australian Organisations Racing To Deploy AI Agents Without Adequate Identity Security Controls, Semperis Study Finds

Australian organisations are granting AI agents access to critical identity systems faster than they are implementing the security guardrails needed to protect them, according to a new global study from identity security firm Semperis.

The State of Identity Security in the AI Era report, based on a survey of 1,100 organisations across eight countries conducted by research firm Censuswide, paints a concerning picture for Australian enterprises as they rush to integrate AI into sensitive identity infrastructure including Active Directory, Entra ID and Okta.

A growing gap between adoption and preparedness

Four in five Australian organisations (80 per cent) believe AI will increase attacks on identity infrastructure, putting them above the global average of 74 per cent. Yet despite this awareness of rising risk, the study found local organisations are pressing ahead with AI deployments that touch some of the most sensitive parts of their IT environments.

Almost a quarter of surveyed Australian organisations (24 per cent) already use AI agents to manage security-related help desk tickets, including password resets and VPN access. Another 69 per cent intend to do so within the next year, bringing the expected total to 93 per cent.

At the same time, 92 per cent of respondents indicated that some portion of their workforce has AI installed on local machines where it can access SSH and encryption keys.

“The accelerated use of AI is introducing a bevy of new agents, each with its own non-human identity throughout global enterprises and many companies are just way too optimistic about their ability to recover their identity infrastructure following a breach, even as they expand this landscape of NHIs,” warned Alex Weinert, Semperis Chief Product Officer.

Recovery confidence well below global average

One of the most striking findings for Australia relates to confidence in recovery. Only 21 per cent of Australian respondents are very confident they could fully regain control of their identity infrastructure if an AI agent exposes admin credentials to an attacker. That figure sits 11 percentage points below the global average of 32 per cent.

One in ten Australian organisations (10 per cent) went further, stating they were not confident in regaining control at all.

The report highlights that non-human identities already vastly outnumber human users in enterprise environments. Microsoft estimates a ratio of roughly 10 NHIs per human identity as of 2018, trending toward 100:1 as agentic and workload identities proliferate.

Each new AI agent, service principal and low-code helper becomes another potential entry point to identity systems. AI support agents are often overpermissioned, with the potential to reconfigure security settings or grant access that can lock teams out of their identity systems or punch holes in corporate VPNs.

“The pattern of global organisations overestimating how quickly they can recover from a cyberattack is real, especially when identity is within the blast radius. On paper, organisations have plans and backups; in practice, identity failures turn technical incidents into prolonged business crises, exposing a dangerous gap between perceived resilience and reality,” observed Chris Inglis, the first U.S. National Cyber Director and Semperis Strategic Advisor.

Governance gaps and identity sprawl

The study revealed significant gaps in how AI identities are being governed in Australia. Only half of Australian organisations (52 per cent) say their AI identities are fully registered, authenticated and authorised in a formal system, compared to 65 per cent globally.

Globally, 6 per cent of organisations admit they do not track AI identities at all. Combined with evidence that most permissions in identity systems are unused or overpermissioned, and that 80 per cent of workload identities are effectively abandoned but still retain access, this creates fertile ground for what the report describes as “zombie” agents and shadow NHIs that attackers can quietly hijack.

In organisations that do track AI identities, 62 per cent of Australian respondents use the same system as for human identities, while 38 per cent authenticate and authorise them using a separate system.

Sarah Cecchetti, Director of Product Management at Semperis, pointed to the challenges of both approaches. Including AI agents as first-class citizens in the identity provider makes sense for using existing roles, permissions and auditing, she explained, but agents may only exist for 30 seconds and their functions may not map to existing roles. The result is that directories can expand to hundreds of times their normal size and agents end up with more entitlements than they actually need.

“CISOs are getting steamrolled without the ability to properly scope and audit these new workloads,” Cecchetti added.

Australia lagging on AI identity governance

Australia recorded the lowest AI governance prioritisation rate of any country surveyed, at 79 per cent. While that figure may appear high in isolation, it trails the global average of 83 per cent and places Australia at the bottom of the eight-country ranking.

Gerry Sillars, Vice President of APJ at Semperis, highlighted the widening gap between Australian organisations and their international peers.

“The data reveals that Australian organisations are lagging behind their international peers when it comes to governing AI-related identities. Locally, organisations are racing to introduce AI identities, despite lacking the visibility and controls needed to securely manage them at scale,” Sillars explained.

“Compared to their global counterparts, Australian organisations also express less confidence in their ability to regain control of their identity systems if AI were to expose their admin credentials. It is clear that AI is changing the identity threat landscape faster than Australian organisations can adapt.”

The attack surface is widening

The report highlights how AI agents sitting on local machines with access to SSH keys, password managers and browser sessions present a direct risk. An attacker who compromises an endpoint or socially engineers the agent could ask it to enumerate credentials and vulnerabilities at machine speed.

Generative search capabilities can traverse overpermissioned resources, potentially summarising active vulnerabilities and known credentials for gaining privileged access into identity systems, Weinert warned. In environments where agents act as users against Active Directory, Entra ID or Okta, or hold API keys and OAuth tokens for Tier 0 systems, a compromised agent can chain those capabilities to impersonate admins, exploit excess permissions and modify policies, groups or conditional access to entrench itself.

Microsoft has reported that identity-based attacks increased by 32 per cent in the first half of 2025, building on what the company described as 600 million or more identity-based attacks per day in 2024.

Country comparisons show mixed preparedness

The study’s country-level data reveals varying levels of preparedness. On the question of whether AI will increase attacks on identity infrastructure, Germany topped the list at 84 per cent, followed by Australia at 80 per cent. France and Italy recorded the lowest levels of concern at 63 per cent and 62 per cent respectively.

On the question of recovery confidence following an AI-related credential exposure, the United States led with 53 per cent expressing high confidence, followed by the United Kingdom at 37 per cent. Australia and Singapore tied for third-lowest at 21 per cent each, with France recording the lowest confidence at just 12 per cent.

For AI identity governance, the United States and Germany both recorded 72 per cent of organisations fully registering, authenticating and authorising AI identities, while Australia sat at 52 per cent and Italy at 69 per cent.

Industry breakdown reveals uneven risk

Industry-level data showed that oil and gas reported the highest concern about AI-driven attacks on identity infrastructure at 100 per cent, followed by utilities at 92 per cent. Government and banking recorded 76 per cent and 80 per cent respectively.

Government organisations recorded by far the lowest recovery confidence at just 4 per cent, while IT and telecoms reported the highest at 35 per cent.

On AI agent use for security help desk tickets, oil and gas reported the lowest current adoption at 11 per cent, but the highest planned adoption, with 89 per cent intending to deploy within the next year.

A call for tighter controls

Grace Cassy, Partner at Ten Eleven Ventures, underlined the urgency of the findings.

“What is striking is not just how quickly AI is being integrated into identity systems but how unprepared many organisations are to recover when things go wrong. Introducing AI at the identity layer offers operational advantages, but it must be accompanied by guardrails, observability, and recovery readiness. It is a new dimension of an old question, really: Are you resilient enough to respond in the event of critical disruption,” Cassy explained.

The report recommends organisations treat agents explicitly as non-human identities in the identity fabric, enforce least-privilege, just-enough and just-in-time access for agents as rigorously as for humans, segregate agent and human trust boundaries, use behavioural analytics to detect anomalous agent behaviour and ensure organisations can quickly recover identity systems to a trustworthy state if breached.

As the report concludes, if organisations are going to let AI touch keys, tickets or Tier 0 systems, they must assume those agents will eventually be compromised and design their identity fabric, backup and recovery and governance controls accordingly.

The full report is available at semperis.com.

Last Updated on May 14, 2026 by Nick Ross