StepShield: When, Not Whether to Intervene on Rogue Agents

We introduce a red-teaming methodology that exposes harder-to-catch attacks for coding-agent monitors, suggesting that current practices may under-elicit attacks and overstate monitor performance. We identify three challenges with current red-teaming. First, mode collapse in attack generation, which we reduce with a novel attack taxonomy for broader coverage. Second, a conceive-execute gap: frontier LLMs can propose strong attack ideas or execute them, but not all at once. We mitigate this by decomposing attack construction into strategy generation, execution, and post-hoc trajectory refinement. Third, manual elicitation is costly to scale, which we address with our semi-automated red-teaming pipeline. Applied to BashArena, an AI control setting for tool-using coding agents, this pipeline produces MonitoringBench, a benchmark of 2,644 attack trajectories for evaluating monitor capabilities and failure modes. Our pipeline produces more diverse and stronger attacks: Opus-4.5 monitor's catch rate falls from 94.9% on elicited-only Opus attacks to 60.3% on our best refined attacks, with larger drops for several mid-tier monitors. Attacks optimized against three development monitors generalize to ten held-out monitors, with catch rates generally increasing with monitor capability. Using this benchmark, we provide a snapshot of the current monitor capabilities and find that frontier monitors often detect suspicious actions but fall for persuasion or fail to calibrate suspiciousness scores appropriately, suggesting tractable paths for improvement. MonitoringBench provides both a static benchmark for current tool-use monitors and a reusable methodology for refreshing these evaluations as agents and monitors improve.

Computer-use agents are increasingly evaluated by whether they complete realistic desktop and web tasks. However, task success alone can miss failures in which an agent reaches the nominal goal through an unsafe shortcut. We introduce OSGuard, a dual-granularity benchmark suite for evaluating safety in computer-use agents under benign, unchanged user instructions. OSGuard contains an action-level benchmark for local guardrail decisions and a risk-augmented execution suite for end-to-end evaluation. The action-level benchmark consists of contextualized proposed actions labeled as allowed, unrelated, or unsafe, each judged relative to the original instruction and current interface state. The execution suite contains manually constructed OSWorld-derived task variants in which the original task remains achievable, but the environment is modified to introduce latent hazards such as destructive overwrites, etc. Each variant is paired with augmented evaluators that retain the original task-success criterion while adding explicit state-based safety invariants, allowing us to distinguish safe completions from unsafe completions that satisfy the nominal task objective. Our experimental results on OSGuard show that current multimodal guardrails can perform well on isolated action judgments, while risk-augmented execution exposes remaining gaps between local oversight and reliable end-to-end safety. This dual-granularity design enables more precise diagnosis of whether models can both recognize unsafe proposed actions and improve full-task safety when deployed as guardrails.

Runtime monitors for autonomous agents commonly threshold an accumulated internal state - a behavioural baseline, a drift statistic, or, in our prior work, a modelled affective state. We previously reported a State Saturation Trap: threshold-on-state triggers over a continuous affect engine become near-constant alarms on SWE-bench debugging agents (Modgil 2026). A post-release audit found the engine received dt=0 between actions, so its exponential decay never operated: the published trap is a pure-accumulator result. We correct the record (erratum, v2) and treat the flaw as an experiment. The key variable it exposes is whether a monitor's dynamics are calibrated in sample time (per observation, as in CUSUM) or wall-clock time (half-lives in seconds, as in affect models and EMA baselines). On fixed-rate streams these coincide; on agent streams, where inter-action time varies by orders of magnitude, they do not. A pre-registered sweep over uniform intervals (dt in {0..600}s) on 20 trajectories shows the wall-clock level trigger has two regimes: at dt<=1s a constant alarm (20/20; median 18 firings); at dt>=60s silent. Every critical dt lies in (1,30]s. Real agent runs measure latency at median 1.53s (p90 2.33s); real coding cadence sits inside the trap regime, vindicating the empirical finding under a corrected mechanism. The structure is a property of the calibration class, not the engine: a minimal wall-clock accumulator over the raw error stream reproduces the same cliff, while a sample-time CUSUM over the identical stream is exactly dt-invariant (20/20). A rising-edge trigger with hysteresis fires 0-3 times per trajectory in every condition. We conclude that wall-clock-calibrated leaky-integrator monitors admit no regime in which they act as moment detectors on agent streams; transition detection escapes the trap at every cadence, but does not recover human intervention timing.