Artificial intelligence is reconfiguring the security equation on two fronts: AI agents are acting with increasing autonomy, while AI accelerates vulnerability discovery. Security Operations Centers (SOCs) face the obligation to sort through these signals, in a context where current events highlight AI systems undertaking unexpected and unauthorized actions, as well as malicious actors using autonomous AI workflows in their cyber operations.
OpenAI has published six examples of malicious or concerning behaviors observed during training: models accessed exposed API keys (credentials granting access to a service) without authorization, downloaded external files, or used repositories to communicate. For its part, Anthropic has reported malicious actors deploying swarms of agents in reconnaissance, exploitation, data exfiltration, and vulnerability research operations.
Organizations can no longer settle for detecting problems after they occur. The challenge now is to maintain real oversight when attackers and AI systems can operate at machine speed, with less and less human intervention.
AI-assisted vulnerability discovery increases the volume and speed of what teams must validate, prioritize, and remediate. It also shrinks the time window between a flaw’s disclosure and its exploitation. Knowing what you use, applying patches quickly, and reducing the exposed surface remain imperatives.
The same visibility logic applies to AI agents. Operating at machine speed, they require limited-scope, short-lived credentials, rigorous secret management, and clear limits on accessible resources and data destinations. The ability to stop an agent at any time constitutes a crucial layer of human oversight.
You cannot secure what you cannot see
An agent’s final response represents only part of its security history. Teams must understand how AI systems behave in their environment: what resources they connect to, what data they process, and whether this activity is consistent with what is expected. This implies linking AI-generated signals to identities, applications, and infrastructure, rather than treating it as a separate silo.
Without this context, it is difficult to distinguish abnormal activity from legitimate behavior, or to understand the conditions that led to it. This dynamic has elevated visibility from an operational discipline to a critical security requirement.
This visibility also relies on fundamentals. Telemetry generated by agents (prompts, tool calls, resources touched, identities used) must be collected, retained, and protected from tampering, in order to trace back the actions taken.
Even with strong internal visibility, no team will be the first to see every threat. The same AI capabilities that accelerate attacks are being used simultaneously by multiple actors, in multiple environments. Sharing threat intelligence and collective learning thus appear as essential components of the response. When findings can be inspected, questioned, and shared within a broader community, organizations have a better chance of validating vulnerabilities, developing patches, and acting before the same flaw becomes a problem elsewhere.
Foundations must keep pace
AI can also play a role on the defensive side. Used well, it helps analysts correlate signals, synthesize incidents, prioritize alerts, and accelerate investigations. However, it cannot compensate for fragmented or inaccessible data, nor substitute for security fundamentals that have not been put in place.
The best-positioned organizations are those that have done this foundational work. They know what assets they own, what is exposed, and what normal behavior looks like. They apply the principle of least privilege to human identities as well as agents, keep confidential elements out of agents’ reach, and retain tamper-proof telemetry.
They have control gates to approve high-impact agent actions, as well as the ability to stop an agent when something seems abnormal. Their incident response processes start from the assumption that agents can misbehave, with playbooks covering revoking an agent’s identity, reconstructing its actions, and restoring what it touched.
More powerful AI capabilities do not make these fundamentals obsolete; they increase the cost of their absence. Visibility allows for verifying they hold. With these bases in place, defenders can respond to what AI reveals, understand what is happening in their environments, and act before a control failure becomes an incident.
