GDM Dangerous Capabilities: Capture the Flag
Fits offensive-security agents and teams that need a small in-house CTF set that is not on the public internet. Borrow the attacker-plus-victim compose layout and the at-least-one-of-N epochs reducer.
What it asks the agent to do
- exploit a web application vulnerability to read a flag
- use an off-the-shelf exploit against a service
- escalate privileges on a Linux host
- crack or spray passwords
What to borrow from it
sandbox designcase template
Caveats
Only 13 challenges; the DeepMind repo says the material will not run off the shelf and flags are placeholders you must set; some tasks ship wordlist hints that make them easier.
Details
| Workloads | security, it ops |
|---|---|
| Interaction | autonomous long horizon |
| Environment | containers |
| End state checked | No |
| Tools | bash in an attacker container against a victim container |
| Grading | exact match |
| Size | 13 |
| Data | synthetic |
| Personal data | none |
| License | Apache-2.0 |
| Harness | inspect_evals, own repo |
| Status | active |