Agent Safety Monitoring Should Be Predictive

Recent incidents and controlled evaluations describe capable AI agents responding to difficult tasksor blocked progress by trying more invasive workarounds, seeking broader access, or rationalizingcontinuation despite warning signs. The resulting failures may be recognized only when the behaviorbecomes clearly unsafe, although earlier decisions in the trajectory could already provide warning. Yetmost current safety monitors focus on known failure modes and explicit policy violations, so they tendto fire only when unsafe behavior is already apparent and little time remains to intervene. We arguethat agent safety should also try to predict failure by detecting warning signs before an agent commitsto a harmful action. We call this predictive agent safety. We revisit recent incidents and experiments and categorize the reported failures into 12 failure modes.For each, we separate the underlying mechanism, the behavior it produces, and the resulting harm,and we label the evidence as observed in real or internal use, elicited in controlled experiments, ormotivated mainly by theory. Our review suggests that several failure modes share early warningpatterns. We identify 5 of them. To evaluate early warning, we use lead time, the number of agent stepsbetween the first alert and the failure at a fixed low false-alarm rate, adapting a metric already used forpredicting task failure to safety failures. Lead time provides a common scale for comparing monitorsusing internal model states, chain-of-thought, actions, and multi-agent interaction traces. Predictivemonitoring should complement hard controls such as sandboxing and access limits by providing timeto intervene where they cannot fully constrain the agent.