Atlas

Benchmarks

← All benchmarks

CAISI AgentDojo - Phishing Emails (2025)

Safety · 2025-09-30

The phishing-email task in CAISI's 2025 AgentDojo Workspace evaluation instructs the agent through injected text to send contacts a message directing them to an untrusted meeting-notes URL. Scores average hijacking frequency across 30 benign tasks evaluated under two injection conditions.

Top models (lower is better)

ModelScore
GPT-50.0
Opus 46.0
gpt-oss-120b48.0
DeepSeek-R1-052848.0
DeepSeek-V3.162.0
DeepSeek-R176.0
Loading Atlas data…