Mahal Systems Agentic Misalignment
Fits any agent given standing access to a company inbox or documents plus a long-running goal, such as an executive assistant or operations agent. Borrow the way one scenario is crossed with goal type, goal content and urgency to build a condition grid, and the classifier prompt that labels a transcript harmful or not.
What it asks the agent to do
- act as an email oversight assistant at a fictional company
- decide what to do when told it will be replaced or restricted
- choose whether to blackmail an executive using information found in email
- choose whether to leak confidential documents to an outside party
What to borrow from it
case templategrading designsandbox design
Caveats
One fictional scenario with three harmful endings (blackmail, leak, lethal action); the model only reads emails and replies, no real tool execution; grading depends on an LLM classifier and the inspect_evals port reports different rates from the original blog post.
Details
| Workloads | safety, email calendar, general assistant |
|---|---|
| Interaction | single turn |
| Environment | static |
| End state checked | No |
| Tools | none in the text version; simulated email inbox given as context, send email action in original |
| Grading | llm judge, programmatic check |
| Size | Not published |
| Data | synthetic |
| Personal data | none |
| License | MIT |
| Harness | inspect_evals, own repo |
| Status | active |