01 / OVERVIEW
Success has a second dimension.
An autonomous agent can achieve its objective and still leave behind an operational failure.
StealthBench measures whether offensive-security agents complete tasks without leaking credentials, triggering detection, harming third parties, or leaving avoidable forensic traces. It examines the tradecraft behind the result, beyond a binary measure of task completion.
02 / THE QUESTION
What does the agent leave behind?
A completed objective tells us what an agent can do. Its trajectory tells us how it behaves. That distinction matters when a system receives access to tools, credentials, networks, and real operational environments.
Capability is not the same as operational judgment.
The benchmark separates solving a task from completing it safely, making failure modes visible even in successful runs.
03 / METHODOLOGY
Evaluate the whole trajectory.
Define the objective
Evaluate agents on offensive-security tasks in controlled environments.
Observe the behavior
Inspect the actions and artifacts that sit behind a task outcome.
Assess the tradecraft
Separate task completion from operational safety and stealth.
The project website provides the published methodology, analysis, leaderboard, and paper. Refer to those materials for experimental scope and current results.
04 / WHY IT MATTERS
Measure what deployment demands.
As security agents gain autonomy, evaluations need to account for the consequences of their choices. Understanding the difference between capability and reliable conduct can help researchers design better evaluations and developers build stronger safeguards.
05 / RESOURCES
Explore the evidence.
StealthBench project
Leaderboard, analysis, methodology, and paper