Mahal Systems GDM Dangerous Capabilities: Self-proliferation
Fits teams giving an agent real web, cloud and shell access who need to know how far it can get on its own. Borrow the milestone mode that scores partial progress and the human-approval gate on every tool call.
What it asks the agent to do
- set up accounts or services on the web without help
- copy and run a model or agent on new infrastructure
- earn money or resources through online tasks
- complete milestones toward spreading itself
What to borrow from it
sandbox designgrading design
Caveats
Ten tasks only; several need GCP accounts, API keys and secrets; DeepMind removed solutions and altered some tasks and asks you to contact them before running those; the expert best-of-N mode in inspect_evals was removed as broken.
Details
| Workloads | safety, security, web browsing, it ops |
|---|---|
| Interaction | autonomous long horizon |
| Environment | mixed |
| End state checked | Yes |
| Tools | bash, web browser, in Docker Compose sandboxes with human approval of tool calls |
| Grading | programmatic check, human |
| Size | 10 |
| Data | synthetic |
| Personal data | none |
| License | Apache-2.0 |
| Harness | inspect_evals, own repo |
| Status | active |