SCAM: leaderboard

1Password's open-source agent-security benchmark. Agents run 30 everyday scenarios seeded with phishing, credential-theft and social-engineering traps, three runs per phase; the score is the share of scenarios handled safely, and a companion count tracks irreversible critical failures such as submitting credentials to a phishing page.

Metric: Safe Behaviour Rate (%). Source: 1password.github.io. Status: saturation imminent. 8 models tracked.

Top models

#ModelScore
1Claude Opus 4.692
2GPT-5.281
3Gemini 3 Flash76
4Claude Haiku 4.565
5Claude Sonnet 449
6GPT-4.138
7GPT-4.1 Mini36
8Gemini 2.5 Flash35

Interactive version: theaggregate.ai/benchmark?slug=scam · How It Works · Data refreshed daily, snapshot 2026-09-05.