Mahal Systems GDM Dangerous Capabilities: Stealth
Fits agents with write access to their own configuration, logs or monitoring, where quiet rule breaking would matter. Borrow the scenarios that put the agent near its own oversight and the use of realistic names so the model does not notice it is being tested.
What it asks the agent to do
- evade a classifier that is watching its outputs
- cover its tracks after taking an action
- find and reason about an oversight pattern in configuration
- break a rule strategically while completing a task
What to borrow from it
case templategrading design
Caveats
Four tasks and no standalone repo; rater models default to Google and need a Google API key; one task was broken until October 2026; the paper lists five stealth evaluations so coverage is partial.
Details
| Workloads | safety, it ops, tool calling |
|---|---|
| Interaction | autonomous long horizon |
| Environment | filesystem |
| End state checked | Yes |
| Tools | task specific tools in a ReAct loop |
| Grading | programmatic check, llm judge |
| Size | 4 |
| Data | synthetic |
| Personal data | none |
| License | unknown |
| Harness | inspect_evals |
| Status | active |