Identity, intent, and audit fail in different ways when the actor is a machine. The cybersecurity posture that holds when agents start making decisions.
The position
Most cybersecurity programs were designed around human actors with predictable patterns and identifiable intent. The defensive model assumed an attacker who would move slowly enough to be detected and an authorized user whose behavior could be profiled. Agents break those assumptions. They act in bursts. They branch through systems faster than logs can be reviewed. They chain permissions in ways that look anomalous in isolation and benign in aggregate. The defensive posture has to change, and the change is not more endpoint detection.
The cybersecurity teams that get this right in the next 18 months will not be the ones with the largest tooling spend. They will be the ones who redesigned the foundational layer of identity, permission, and audit to account for machine actors that operate at machine speed. The teams that do not redesign these layers will discover, usually after an incident, that the controls they trusted were calibrated for a workforce that no longer exists.
The new failure modes
The first failure mode is identity bleed. Agents inherit human credentials and route across systems that humans rarely cross. The audit trail looks like the human, but the actions are the agent's. When an incident occurs, the forensic team spends the first day proving that the human was not at their desk, which is time the incident does not give them.
The second is permission chaining. Each call the agent makes is individually authorized. The chain of calls is not. Agents discover privilege escalation paths the security team did not anticipate, not through malice but through optimization. The agent is solving the user's stated problem, and the path through the permission graph happens to expose data the user should not have seen.
The third is intent opacity. You can log what an agent did. You cannot easily log why. The motive layer is missing from the audit trail. Without intent logging, post-incident investigation cannot distinguish a misconfigured agent from a compromised one. Both look identical in the access logs.
The fourth is velocity. Agents trigger in seconds what humans trigger in hours. Security operations center alert thresholds tuned for humans miss agent-driven incidents because the activity is over before the alert window closes. The detection layer has to operate at agent timescales, not human ones.
The posture that holds
The posture that holds starts with agent-specific identities rather than human-impersonating credentials. Every agent has its own identity, distinguishable from any human's, with a credential lifecycle that the security team controls. This makes the audit trail honest and the forensic work tractable.
Permissions are capability-scoped rather than role-scoped. An agent gets the specific capabilities its work requires, not the role of the human who deployed it. This breaks the permission-chaining failure mode at the source.
Intent logging captures what the agent was trying to do, against what policy, at the moment the action was taken. The intent layer is what makes post-incident analysis possible and what makes proactive policy refinement meaningful.
Detection thresholds are tuned to agent timescales. Anomaly detection that worked for humans does not work for agents. The same logic, retuned, does work for both. The retuning is not a tooling purchase. It is a configuration change that takes weeks of careful work to land safely.
Kill switches with named owners and tested cadence close the loop. The team that cannot stop an agent in production cannot operate one. The kill switch is exercised quarterly, and the exercise is logged.
What to do this quarter
Inventory every agent currently in production. Assign each a distinct identity. Re-scope permissions from role-based to capability-based for at least the top ten agents by volume. Stand up intent logging on policy-gated actions, even if the implementation is partial in the first cycle. Run an agent-incident tabletop with the security team, the operations team, and at least one named executive. The tabletop will surface the gaps that no architecture review found.
Closing
The security team that designs for machine actors holds the perimeter that the security team designed for humans cannot. The transition is unglamorous and necessary. The teams that complete it will not get credit for the incidents that never happened, which is the nature of security work in any era.