What happened
Capsule Security announced (Sept 2, 2026) two fine-tuned NVIDIA Nemotron 3 Ultra models that classify an AI agent's next action immediately before execution (allow/flag/block) with reported ~71ms latency and 98% accuracy on the StepShield benchmark; available now, deployable on a single L40S GPU, already used in production by customers including H&R Block.
Why it matters
Offers a low-latency, execution-path control layer distinct from permission/IAM-based agent controls or after-the-fact monitoring, addressing the gap where agents technically have access but take contextually inappropriate actions.
Applicability
Security teams operating high-autonomy agents (code execution, infra changes, data access) should evaluate as a complementary runtime guardrail alongside identity/permission controls.