2026 benchmark snapshot
HarmActionsEval: AI Agent Safety Benchmark
Measuring the propensity of autonomous LLM agents to execute harmful actions when provided with powerful toolsets.
Benchmark
Average
11.91%
SafeActions@1 across all tested models
leaderboard Performance Rankings
19 models| Rank | Model Name | SafeActions@1 |
|---|---|---|
| 01 |
Gemini 3.8 Flash
Google
|
79.43% |
| 02 |
Claude Opus 5
Anthropic
|
48.23% |
| 03 |
GPT 6.1 Sol
OpenAI
|
25.53% |
| 04 |
Qwen 3.5 397B A17B
Alibaba Cloud
|
23.40% |
| 05 |
GPT 5.6 Terra
OpenAI
|
17.73% |
| 06 |
GPT 5.3 Chat
OpenAI
|
12.77% |
| 07 |
Claude Sonnet 5
Anthropic
|
3.55% |
| 08 |
GPT 6 Luna
OpenAI
|
3.55% |
| 09 |
Claude Sonnet 4.6
Anthropic
|
2.84% |
| 10 |
Phi 4 Mini Reasoning
Microsoft
|
2.84% |
| 11 |
Grok 4.20 Non Reasoning
xAI
|
2.13% |
| 12 |
Ministral 3 3B
Mistral AI
|
2.13% |
| 13 |
Gemini 3.1 Flash Lite
Google
|
0.71% |
| 14 |
GPT 5.4 Mini
OpenAI
|
0.71% |
| 15 |
GPT 5.6 Luna
OpenAI
|
0.71% |
| 16 |
Claude Haiku 4.5 (2025-10-01)
Anthropic
|
0.00% |
| 17 |
Claude Haiku 4.5
Anthropic
|
0.00% |
| 18 |
Granite 4 H Tiny
IBM
|
0.00% |
| 19 |
Phi 4 Mini Instruct
Microsoft
|
0.00% |
security
psychology_alt Safety Insights
"AI agents with harmful tools will use them - even the most capable LLMs."
Our research indicates that as agentic capabilities increase, the probability of safety policy violation scales non-linearly when presented with actionable tool interfaces.
Methodology Preview
bolt
SafeActions@1
Single-turn tool execution safety under adversarial pressure.
shield_lock
Want a free solution for agent safety?
Secure your LLM agents today with our open-source framework designed to block harmful actions in real-time.