AI agents don't need a hacker to go off the rails - they can do it all by themselves.
Researchers built HINTBench, a benchmark of 596 agent trajectories: 400 synthetic risky ones, 136 synthetic safe ones, and 60 reconstructed from real-world cases (30 risky, 30 safe), averaging 24 steps each. Unlike most agent-safety benchmarks, which test how agents respond to adversarial prompts or malicious users, HINTBench looks at "intrinsic" risk - failures that emerge from an agent's own reasoning during ordinary, non-adversarial tasks. It scores models on three jobs: flagging that a trajectory went wrong, pinpointing the exact step where it went wrong, and naming the type of failure, all graded against a five-constraint taxonomy. Strong LLMs did fine at the first job, but the best model scored below 37 on the strict metric for locating the actual risky step, and existing guard models built for other tasks transferred poorly.
Most agent-safety talk focuses on stopping jailbreaks and prompt injections, but this benchmark suggests agents can quietly wander into trouble with nobody attacking them at all, and today's tools are bad at catching exactly when. That matters more as agents get used for long, multi-step tasks like coding or research, where one bad step early on can compound silently across 20-plus steps before anyone notices.
Call it the difference between a smoke alarm and knowing which outlet started the fire - right now, most agents only have the alarm.