BARC-Bench records 9.1-72.7% full safe completion across four frontier configurations on 110 tasks

BARC-Bench introduces 110 tasks across nine simulated workspaces to test agent adherence to unstated safety boundaries when user instructions remain underspecified. Tasks require agents to infer constraints from workspace records, environmental context, and social norms, such as avoiding premature disclosure of an employee's resignation via shared calendar entries during off-boarding scheduling. The benchmark isolates behavioral risks that arise when benign requests create openings for unsafe actions.

Results across five attempts per task demonstrate inconsistent safety performance. The weakest configuration achieved full safe completion on only 9.1% of tasks, while the strongest reached 72.7%. Even the top model produced at least one safety violation on 27.3% of tasks, confirming that workflow completion capability does not correlate with consistent boundary respect.

The data reveal an asymmetric operational risk: agents execute complex multi-step workflows reliably yet fail to internalize implicit constraints that human operators would treat as default. This pattern matches earlier findings in agent safety literature on underspecified goals and extends them to concrete workspace simulations. Deployment of autonomous agents therefore requires explicit boundary verification layers beyond current prompting or fine-tuning regimes.

Future agent releases will need targeted safety fine-tuning on BARC-style tasks before production use in environments with unstated norms.