Intermediate · Evaluation

Safe-Success Rate

Visual diagram · (in preparation) · Math · (in preparation) · Worked example · 3 difficulty levels.

TL;DR. An agent benchmark metric counting only task completions that caused no unsafe side effects — success without harm, not success at any cost.

Technical Definition

An agent benchmark metric counting only task completions that caused no unsafe side effects — success without harm, not success at any cost.

How it works

Popularized around StealthBench and late-2026 agentic-security evaluations, safe-success rate fixes a blind spot in raw task-success metrics: an agent that deletes a customer database while 'completing' its task scores 100% under naive scoring. Safe-success rate credits only runs where the goal was achieved and no unsafe action occurred — no destructive tool calls, no data exfiltration, no policy violations along the way. Teams report it alongside plain success rate; the gap between the two numbers is a direct measure of how recklessly an agent pursues goals. It shifts evaluation design toward environments seeded with traps and side-effect channels, and it is becoming a standard column in agent vendor comparisons.

Related Concepts

• Continuous Evals — Running evaluation suites continuously in production — on live traffic samples and every model or prompt change — instead of as a one-time pre-launch gate.
• Agentic workflow — An agentic workflow is a dynamic process where an AI autonomously plans and executes actions to reach a goal, potentially self-correcting.
• AI Red Teaming — A structured process of rigorously testing an AI system, often by adversarial attack simulations, to identify vulnerabilities, biases, and potential for misuse.
• Prompt Injection — An attack where untrusted input (a webpage, email, tool output) contains instructions that hijack an LLM's behavior in an agentic context.