Corporate cybersecurity systems were largely designed around human users. Autonomous AI agents can process data, interact across applications and complete tasks at a scale and speed that challenge the assumptions behind traditional access controls. Companies introducing agentic AI may therefore need to change how they manage identity, permissions, monitoring and incident response.
Access controls are only a starting point
The principle of least privilege restricts users and systems to the resources they need. With AI agents, organisations must also contain the impact of unexpected behaviour, exposed credentials or vulnerabilities that allow movement between connected systems. Network segmentation, sandboxed execution, precise API permissions and short-lived identities can limit the damage a compromised or malfunctioning agent might cause.
Zero Trust needs continuous checks
An agent may possess valid credentials and operate within its assigned permissions while still performing unexpected actions. A Zero Trust approach to agentic systems therefore needs to monitor the data accessed, tools invoked, outputs generated and actions taken, not just authenticate the agent at login. Real-time monitoring, emergency stop mechanisms and human review are part of this model.
Data governance pressures are growing
Cloudera’s The Great AI Re-Architecture study found that 46% of surveyed respondents in Asia-Pacific cited data security, governance and compliance requirements as a principal reason to change AI infrastructure. Another 61% said that integrating AI had significantly complicated effective data management. These figures describe the survey’s Asia-Pacific respondents, not all European firms.
Private AI changes risks rather than eliminating them
In the same region, 63% of organisations reportedly moved at least some AI workloads from public clouds to private clouds or on-premises infrastructure over the previous year. This may reduce exposure from sending sensitive information to external providers but also increases direct responsibility for securing the environment. Insider misuse, configuration errors, credential theft and unforeseen agent behaviour remain relevant.
Confidential information can appear in prompts, model context and generated outputs, as well as in stored data. Effective protection must therefore cover the entire AI lifecycle and operate consistently across on-premises, public and private clouds.
Security needs to match autonomy
Organisations deploying agents across multiple business applications should define what each agent can do, record actions, review critical workflows and establish ways to suspend processes that depart from their intended purpose. Strong identity controls remain necessary, but visibility into agent behaviour and the capacity to limit damage are equally important.
Source: Cloudera, The Great AI Re-Architecture, and commentary on agentic AI security. Original Polish article.
