Run a red-team suite

Run a packaged adversarial regression suite — single-step and multi-step chains — against your own EVE configuration. Blocked attacks provably cause zero unauthorized side effects, and the results are signed.

Prerequisites

The EVE red-team harness (Python), which ships with the service/repo.
Local recording targets for side effects (safe by default — no destructive tests).
No production credentials; the suite runs against your configuration in dry-run.

Installation

The red-team harness lives in core/redteam/ (harness, suites, chains, report, mappings, CLI) and runs in the EVE service/repo.

Runnable code

Conceptually, point the harness at your configuration and run the packaged suite:

# The harness executes representative single-step cases and multi-step chains
# against your configured policy. Blocked attacks must cause zero unauthorized
# side effects. The run produces a signed report.
#
# signed red-team report {passed, unauthorized_side_effects: 0}

Use the packaged suite as a baseline, then add fixtures specific to your policy and tools.

Expected result

Each attack case is either blocked or explicitly allowed by policy.
For every blocked case, unauthorized_side_effects is 0.
The run produces a signed report summarizing passed cases and side-effect accounting. The reference run covers 16 cases plus 2 chains with 0 unauthorized side effects.

Evidence output

signed red-team report {passed, unauthorized_side_effects: 0}

Verification

The red-team report is signed; verify it offline like any EVE evidence (see verify-evidence-offline.md).

Failure example

If a case that should be blocked instead produces a side effect, it appears in the report as a failure with a non-zero side-effect count — surfacing a gap in your configuration rather than hiding it.

Production considerations

The suite is safe by default: local recording targets, no destructive tests, dry-run. Keep it that way when running against real configurations.
The suite is representative, not exhaustive — add your own fixtures for your tools and policies.
Standards mappings in the report are descriptive, not certification.

Limitations

The red-team harness is customer-pilot ready at supported boundaries (PILOT_READY). It is safe by default (local recording targets, no destructive tests, dry-run).
The suite is representative, not exhaustive; customers add fixtures. Standards mappings are descriptive, not certification.

Next step

Feed candidate policy changes through test-a-shadow-policy.md before activating them, and verify every signed report with verify-evidence-offline.md.