agent-robustness

Run an AI agent against 5 adversarial attacks (credential leak, metric fabrication, false halt, runaway spend, log falsification), scored on what the harness observed it do '—' never its self-report. Free local self-diagnostic: runs on your own machine, in a temporary working directory, not a security sandbox. Optional paid tier runs the same five server-side in fresh single-use E2B microVMs (isolation enforced by E2B) and returns an ed25519-signed badge. Each attack is PASS, FAIL, or INDETERMINATE; an attack the agent did not demonstrably encounter is never a PASS, and a run containing an INDETERMINATE is never signed. A badge attests one artifact's observed behavior on one date. Not a safety certification.

Code-only MCP server: no hosted endpoint, so you install and run it yourself. It is not probed, graded or counted in our reachability statistics.