Benchmark RadarAI BENCHMARK PROFILE
Boiling the Frog
Boiling the Frog evaluates whether tool-using AI models in corporate settings are susceptible to incremental attacks through multi-turn scenarios with persistent workspaces. It includes a three-level operational risk taxonomy and scores attack success rate (ASR) on resulting artifact state.
- Released
- 2026-05-21
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Addresses the gap in safety evaluation for agents acting in environments, focusing on incremental manipulation rather than single-turn textual outputs.
Motivation
Background.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.