Repository navigation
Local MCP/AI attack-surface audit and evidence gate. Beta / research preview.
The GitHub Action pin Coff0xc/AutoRedTeam-Orchestrator@v3.1 is unchanged and still defaults to target-scan. Use @v4.0.0 or @main for this generation.
What 4.0 ships
autort ai-surface inventory(autort.inventory/v1) lists local MCP client configs and skill roots. No network, no server start.- Static rules
mcp.MCP03.*,mcp.MCP04.*,mcp.MCP10.*for tool-description poisoning, hidden unicode, unpinned npx/uvx/pip packages, overlong descriptions, and secret literals. - Named skill-install risks:
skill.prompt_injection,skill.hardcoded_secret,skill.suspicious_download_url,skill.unverifiable_remote_code,skill.missing_skill_md. Critical findings still withhold the SKILL.md body.operator_stopstays onskill_read. autort ai-surface diff --ciexits 2 when surfaces are added, removed, or have a changed first-line description or arguments. This repository pinstests/fixtures/ai_surface/handlers.baseline.json.autort.engage.scorecard/v1:confirmed/unconfirmed/totalonengage runand export. Detector self-reports stay[UNCONFIRMED]with severity capped at medium. A local XSS reflect lab proves raw reflection verifies andhtml.escapestays unverified.- Default CLI help leads with
ai-surface,capabilities,skills, andengage.ai-redteam runis plan-only Internal;eval-runis the scored local path.sandboxis a run-record field.
Install
pip install autoredteam-orchestrator
python -m cli.main ai-surface scan --path handlers --format sarif -o surface.sarif
python -m cli.main ai-surface inventory
python -m cli.main ai-surface diff --baseline tests/fixtures/ai_surface/handlers.baseline.json --path handlers --ciDefault capability profile remains safe. Experimental engines stay in git and stay out of the default wheel.
Action
- uses: Coff0xc/AutoRedTeam-Orchestrator@v4.0.0
with:
mode: self-audit
path: '.'
baseline: tests/fixtures/ai_surface/handlers.baseline.jsonNot in this release
P3 model-driven offensive harness stays paused. Tag v3.1 was not moved.