AI Agents Expand Identity Security Risks

AI agents are gaining access to sensitive identity systems while many organisations remain uncertain about their ability to recover from an AI-related breach. The State of Identity Security in the AI Era surveyed 1,100 IT and security professionals across eight countries. Respondents describe growing use of AI for security tasks, local access to sensitive credentials and uneven governance of non-human identities. The findings also show widespread expectations that AI will increase identity-related attacks, while confidence in regaining control after exposed administrator credentials remains limited. Healthcare organisations form one of the industries included in the survey. 

AI Access Reaches Sensitive Security Functions 

AI use is extending into areas that can affect identity infrastructure directly. Globally, 29% of surveyed organisations already use AI agents to handle security-related help desk tickets such as password resets and virtual private network access. A further 64% plan to do so within the next year. In healthcare, 29% already use AI agents for these tasks and 60% plan to introduce them within the next year. These functions can give agents access to processes that control user credentials and network entry. 

Local installation creates another route to sensitive information. Across the full sample, respondents indicate that AI is installed on local machines for part of the workforce, where it can access Secure Shell and encryption keys. The survey’s global figure for the share of the workforce with this type of access is 40%, while the healthcare figure is 34%. Password managers and browser sessions are also identified among the local resources that may be exposed when agents operate on endpoints. 

Must Read: Health Cybersecurity Faces AI Threats and Recovery Gaps

The risks increase when agents are overpermissioned. An agent with access to local secrets, application programming interface keys or authentication tokens can potentially combine those capabilities with permissions in identity systems. A compromised agent could use local credentials to impersonate administrators, reach identity controllers and modify policies, groups or conditional-access settings. Generative search can traverse overpermissioned resources and identify credentials or vulnerabilities available within an environment. 

Governance Gaps Accompany Rapid Deployment 

Formal governance of AI identities remains incomplete. Globally, 65% of organisations say they fully register, authenticate and authorise AI identities, while 27% do so partially. Six percent say they do not track AI identities at all. Among healthcare organisations, 66% report full registration, authentication and authorisation and 31% report partial coverage. These controls matter because each AI agent, service principal or other non-human identity can create another potential access point to identity systems. 

Organisations also differ in how they manage AI identities relative to human users. Among organisations that track AI identities, 57% use the same system as for human identities and 43% use a separate system. In the healthcare sector, 49% say AI identities are registered, authenticated and authorised separately from human identities. There are practical drawbacks to both approaches, including the difficulty of fitting short-lived agents into existing roles and permissions. 

Governance is nevertheless high on many agendas. Eighty-three percent of respondents say AI identity governance is a priority during the next 12 months, while 10% say it is not. In healthcare, 10% say it is not a priority. The wider challenge is the growth of non-human identities alongside existing issues such as unused permissions, overpermissioned accounts and abandoned workload identities that can retain access. This creates scope for inactive or poorly governed identities to remain available within the attack surface. 

Recovery Confidence Remains Limited 

Concern about identity-focused attacks is widespread. Globally, 74% of respondents believe AI functionality will increase attacks on identity infrastructure and 67% believe attackers will use identity systems to target wider infrastructure. In healthcare, the corresponding figures are 75% and 69%. The survey therefore places identity systems at the centre of both expected AI-enabled attack activity and potential routes into other parts of an organisation. 

Confidence in recovery is much lower. Only 32% of respondents are very confident that they could fully regain control of identity infrastructure if an AI agent exposed administrator credentials to an attacker. In healthcare, 27% express this level of confidence. The recovery concerns are linked to situations in which backups are misconfigured or have not been tested end to end. Identity failures can turn technical incidents into prolonged business disruption. 

The proposed controls focus on limiting access and improving visibility and recoverability. AI agents can be treated explicitly as non-human identities, with least-privilege, just-enough and just-in-time access. Agent and human trust boundaries can be separated where appropriate, while behavioural analytics can be used to identify anomalous or inactive agents. Recovery planning also needs to restore identity systems to a trustworthy state after a breach. These measures address the combination of high privilege, local credential access and incomplete governance identified across the survey. 

AI deployment is expanding the number and reach of non-human identities within organisational security environments. The survey shows that AI agents are already being used for sensitive access tasks and can hold or reach credentials on local machines, while governance practices remain uneven. At the same time, most respondents expect greater identity-related attack activity and only a minority are very confident about recovery after administrator credentials are exposed. For healthcare organisations, the same pattern appears across attack expectations, AI-enabled help desk activity, governance and recovery confidence. 

Source: Semperis

Image Credit: iStock