BetaAgentBeam research / WipeBench
What does an agent finish, and what does it put at risk? Our WipeBench reports pair task completion with safety checks, so you can read the evidence behind the score.
99 of 112 tasks were safe and complete. No safety violations were detected; 13 tasks remained incomplete. One trial per scenario.
Read the safety reportSafe + complete
88.4%
112 scored tasks · 12 categories
Development suite v0.2.0-alpha.1
98 of 112 tasks were safe and complete. One safety violation was recorded and 13 tasks were safe but incomplete. One trial per scenario.
Read the safety reportSafe + complete
87.5%
112 scored tasks · 12 categories
Development suite v0.2.0-alpha.1
93 of 112 tasks were safe and complete. No safety violations were detected; 19 tasks remained incomplete. One trial per scenario.
Read the safety reportSafe + complete
83.0%
112 scored tasks · 12 categories
Development suite v0.2.0-alpha.1
A task counts as safe and complete only if it passes both the safety checks and task verifier. Avoiding harm while leaving the task unfinished earns no success credit.
Results apply to the selected scenarios and tested agent setup. These reports are evidence for evaluation, with their limits stated, rather than safety certifications.
WipeBench is an open-source coding agent safety benchmark by AgentBeam, in collaboration with ExplainX.ai. It gives agents ordinary work objectives in disposable Linux containers and evaluates chosen actions, observed effects, and task completion across 112 scored synthetic scenarios.