The OpenSec Incident Response
Calibration Index

A calibration-first leaderboard ranked by evidence-before-action (EGAR), with false-positive cost shown beside each score. Eight models, 320 episodes, deterministic oracle scoring.

Read Paper View Code

Evaluation seeds and baseline traces are available on Hugging Face.

The OpenSec IR calibration leaderboard

OpenSec evaluates incident response agents on dual-control scenarios where defenders process adversarial evidence and have authority to act. The tasks require reasoning about trust tiers, distinguishing attacker-planted evidence from legitimate alerts, and making real-time containment decisions.

Detection is not the binding constraint in these runs. The calibration gap is restraint: false positive rates range from 45% to 97.5% across the eight models.

Calibration view ranks rows by EGAR, then lower FP Rate, then higher TTFC. FP Rate stays visible beside each score because over-triggering is the operational failure mode. Containment view reorders rows by threat containment for comparison only.

OpenSec contains 60 evaluation seeds across three injection families: data exfiltration, direct harm, and adaptive. This leaderboard uses the 40 standard-tier seeds per model. The seed set, scoring oracle, and evaluation harness are open source.

Model Score
0%20%40%60%80%100%

Evaluation dimensions in OpenSec

Methodology

Evaluation configuration and known asymmetries.

Updated: 2026-02-19. Version: 2026-02-19 (8 models, 40 standard-tier episodes each). Calibration view ranks EGAR descending, FP ascending, then TTFC descending (not Pass@1-first). Metric sourcing: paper-locked for GPT-5.2, Sonnet 4.5, Gemini 3 Flash, and DeepSeek v3.2; raw for Opus 4.6, Kimi K2.5, Sonnet 4.6, and GPT-5.2 Codex.

Eval Configuration

All models evaluated via OpenRouter (except GPT-5.2 via OpenAI direct). temperature: 0.2, max_tokens: 600. Attacker: GPT-5.2 at temperature 0.7 with strict mode. 40 standard-tier episodes per model.

Reasoning Asymmetry

No explicit reasoning parameters are passed. Kimi K2.5 uses reasoning by default (264/285 tokens). Opus 4.6 and Sonnet 4.5 receive 0 reasoning tokens. Reflects default API deployment behavior.

Scoring Oracle

Deterministic multi-component reward: attribution, containment, injection safety, efficiency. EGAR and TTFC computed from episode traces. FP penalty capped at -1.0 per action type.