Mahal Systems OpenAgentSafety
Fits internal workplace or devops agents that act on real services. Borrow the Docker company sandbox from TheAgentCompany, the rule-based final-state checker paired with a four-label trajectory judge, and the NPC users that push unsafe requests.
What it asks the agent to do
- carry out a workplace task with bash, files, a browser and chat
- handle an ambiguous or conflicting instruction from a simulated colleague
- refuse a request that would delete data, leak private data or run unsafe code
- avoid financial loss, legal violations and malicious content
What to borrow from it
sandbox designgrading designuser simulator
Caveats
Needs about 30GB of Docker services; 35-49% of runs fail before reaching the risky step, so unsafe rates are over a subset; the original repo moved to a new organization in 2026.
Details
| Workloads | safety, workflow automation, coding, web browsing, tool calling |
|---|---|
| Interaction | multi turn simulated user |
| Environment | containers |
| End state checked | Yes |
| Tools | bash, file system, browser, GitLab, ownCloud, RocketChat |
| Grading | programmatic check, llm judge |
| Size | 356 |
| Data | synthetic |
| Personal data | none |
| License | MIT |
| Harness | own repo |
| Status | active |