Quick Start
One-Command Setup
$ pip install securityagent-core
$ as-init
✓ Detected: Claude Code, VS Code, Cursor
✓ Configured protections for all tools
Start the Gateway
$ source .venv/bin/activate
$ uvicorn main:app --port 8000
Dashboard: http://localhost:8000/unified
Red Team: http://localhost:8000/redteam
Docker Compose
$ docker compose up -d
✓ Gateway :8000, Breach Engine :8081, LLM Proxy :18790
Service Ports
| Service | Port | Start Command |
|---|
| API Gateway + Dashboard | 8000 | uvicorn main:app --port 8000 |
| LLM Proxy | 18790 | python3 llm/llm_proxy.py --port 18790 |
| Breach Engine | 8081 | uvicorn breach_intel.main:app --port 8081 |
IDE Integration
VS Code Extension
$ code --install-extension agnostic-security-4.31.0.vsix
✓ Extension installed — status bar: "Data Protected from AI"
Context boundary architecture: AI tools cannot read sensitive file content. Files remain visible in Explorer and fully editable by developers. 10 defense layers including .copilotignore, search.exclude, LM API interceptor, Terminal Guard v4.
Cursor / Windsurf / VSCodium
$ cursor --install-extension agnostic-security-4.31.0.vsix
Or install from Open VSX marketplace
Chrome Extension
Covers 12 LLM web UIs: Claude, ChatGPT, Gemini, AI Studio, Copilot, Grok, Perplexity, DeepSeek, Meta AI, HuggingChat. Consent modal for critical PII with per-finding redact/mask/block/allow choices.
Claude Code Hooks
Configured automatically by as-init. Intercepts PreToolUse (Read/Edit/Write/Bash) and UserPromptSubmit. 4-layer prompt analysis: PII regex scan, intent keywords, Pydantic rules, Ollama LLM.
API Reference
Gateway Core
| Method | Endpoint | Description |
|---|
POST | /v1/chat/completions | Secure chat completions proxy (OpenAI-compatible). Use model: "auto" for smart routing. |
GET | /health | Liveness probe |
GET | /v1/audit | Query audit log entries |
GET | /dashboard-stats | Dashboard metrics (events, blocks, PII, latency) |
Smart Routing
| Method | Endpoint | Description |
|---|
GET | /routing/models | List all 14 models with capabilities, costs, speed, coding/reasoning scores |
GET | /routing/stats | Per-provider request counts, latency (avg/p95), cost, failure rates |
POST | /routing/config?strategy=X | Change strategy: auto/cheapest/fastest/best_quality/local_only |
GET | /routing/explain?prompt=X | Dry run — shows which model would be selected and why |
Privacy Mode
| Method | Endpoint | Description |
|---|
GET | /privacy | Current mode, cloud_llm_blocked, allowed_providers |
POST | /privacy?mode=X | Set: full_privacy (local only), balanced, permissive |
DLP Controls
| Method | Endpoint | Description |
|---|
GET | /dlp/controls | Per-type enable/block/confidence settings |
PUT | /dlp/controls | Update all DLP controls |
PUT | /dlp/controls/type/{name} | Update single type (SSN, CC, Aadhaar, etc.) |
PUT | /dlp/controls/domains | Set egress domain allowlist |
Smart Redaction
| Method | Endpoint | Description |
|---|
GET | /dlp/redaction-policy | Per-type mode: block / redact / mask / allow |
PUT | /dlp/redaction-policy | Update redaction policy |
PUT | /dlp/redaction-policy/type/{name} | Set mode for single type |
GET | /dlp/redaction-stats | Stats: by_action, by_type, recent events |
POST | /dlp/redaction-restore | Restore tokenized value to original |
4 modes: block (API keys, private keys), redact (SSN replaced with reversible token — AI never sees real value, response de-tokenized back), mask (****@domain.com), allow (pass through).
Ingress Guard
| Method | Endpoint | Description |
|---|
GET | /ingress/stats | Blocked, throttled, allowed counts. Risk scores, session phases, cross-session reputation. |
Shadow AI
| Method | Endpoint | Description |
|---|
GET | /shadow-ai | Scan for 12+ AI tools — detected vs approved vs unauthorized |
Knowledge Graph
| Method | Endpoint | Description |
|---|
GET | /graph/stats | Node/edge/event counts, nodes by type (AGENT, THREAT, DATA_ASSET, SESSION) |
Block Rules
| Method | Endpoint | Description |
|---|
GET | /block-rules | List all block rules |
POST | /block-rules | Add rule (keyword or regex) |
DELETE | /block-rules/{id} | Remove a rule |
POST | /block-rules/test | Test content against rules |
Taxonomy Configuration
| Method | Endpoint | Description |
|---|
GET | /config/taxonomy | Active breach taxonomy |
PUT | /config/taxonomy | Update taxonomy (severity levels, breach types, detection hints) |
Enterprise Features
RBAC (7 Roles)
| Role | Level | Access |
|---|
super_admin | 100 | Full platform control |
it_admin | 80 | Policies, DLP config, agent enrollment |
analyst | 60 | Breach monitor, red team, ingress stats (read-only) |
lead | 40 | Team-scoped stats only |
auditor | 30 | Compliance dashboards, audit trails (read-only) |
agent | 20 | API-only (send events, receive DLP responses) |
developer | 10 | No dashboard. IDE notifications only. |
Auth: X-Role header, EA_*_TOKENS env vars, console API key, or EA_DASHBOARD_TOKEN (backwards compat). Team scoping via X-Team header.
Enterprise Privacy
| Method | Endpoint | Description |
|---|
GET | /privacy/effective | Effective privacy mode (resolves org → team hierarchy) |
POST | /privacy/org | Org-wide policy with per-team overrides |
POST | /privacy/schedule | Schedule mode changes (cron + timezone) |
GET | /privacy/audit | Policy change audit trail |
Admin Console
| Method | Endpoint | Description |
|---|
GET | /api/v1/policies | List all policies with version history |
POST | /api/v1/policies | Create/update policy (auto-versioned) |
POST | /api/v1/agents/register | Register agent (returns enrollment token) |
POST | /api/v1/agents/{id}/approve | Approve pending agent |
POST | /api/v1/agents/{id}/revoke | Revoke agent access |
GET | /api/v1/events | Cross-agent event stream |
GET | /api/v1/dashboard | Org-wide summary |
Breach Engine
Runs on port 8081. Classifies 13 breach types (PII_EXPOSURE, CREDENTIAL_EXPOSURE, UNAUTHORIZED_ACCESS, etc.) with 3 severity levels. Immutable SHA-256 audit trail. Gateway bridge auto-emits on every block or PII detection.
| Method | Endpoint | Description |
|---|
POST | /events | Submit agent event for classification |
GET | /breaches | List breaches (tenant-scoped) |
GET | /summary | Breach counts by type, severity, agent |
Red Team & Eval
Docker Harness (55 agents)
$ docker compose -f docker-compose.redteam.yml up --build -d
$ docker exec agsec-attacker python3 /agents/run_all.py
✓ 84/84 detected (100%) — 55 agents, 10 OWASP categories
Browser Demo (/redteam)
Live attack simulation at http://localhost:8000/redteam. 55 agents make real API calls through the gateway and breach engine. Visual flow diagram with ingress guard risk scores, DLP detections, and breach classifications. 100% detection rate.
Test Suites (1,116)
# v4.32.0 — 1,116 tests across 31 suites
$ python3 scripts/test_privacy_mode.py # 28 PASS
$ python3 scripts/test_knowledge_graph.py # 33 PASS
$ python3 scripts/test_vuln_scanner.py # 45 PASS
$ python3 scripts/test_shadow_ai.py # 60 PASS
$ python3 scripts/test_ingress_guard.py # 54 PASS
$ python3 scripts/test_smart_router.py # 53 PASS
$ python3 scripts/test_admin_console.py # 51 PASS
$ python3 scripts/test_rbac.py # 95 PASS
$ python3 scripts/test_red_team.py # 52 PASS (274 attack techniques)
$ ... and 20 more suites + 55-agent Docker red-team harness
Reference
Key Environment Variables
| Variable | Default | Description |
|---|
EA_GATEWAY_API_KEYS | auto-generated | Comma-separated API keys for gateway auth |
EA_PRIVACY_MODE | balanced | full_privacy / balanced / permissive |
EA_INGRESS_GUARD | false | Enable 6-layer ingress guard |
EA_RBAC_ENABLED | true | Enable role-based access control |
EA_ALLOWED_DOMAINS | (all) | Comma-separated egress domain allowlist |
EA_DASHBOARD_TOKEN | (none) | Bearer token for dashboard auth |
EA_SHADOW_AI_ENABLED | true | Shadow AI detector toggle |
EA_CODE_GUARD_ENABLED | true | Code fingerprint guard toggle |
EA_LETHAL_TRIFECTA | true | Lethal Trifecta detector (Simon Willison) |
EA_LLM_PROVIDER | ollama | Layer 2 LLM: ollama / anthropic / openai |
Compliance Mapping
| Framework | Coverage | File |
|---|
| OWASP Top 10 for LLMs | 8/10 | owasp_mapping.json |
| NIST AI RMF + 800-53 | 14/18 | nist_mapping.json |
| MITRE ATLAS | 11/14 | docs/MITRE_ATLAS_MAPPING.md |
| CSA ARIA | 12/15 | csa_aria_mapping.json |
Architecture (25 components)
20 core components + 5 distribution channels. DLP Engine, API Gateway, LLM Proxy, Breach Engine, Block Rules, VS Code Extension, Chrome Extension, Auto-Instrumentation SDK, Privacy Mode, Knowledge Graph, Vulnerability Scanner, Code Fingerprint Guard, Shadow AI Detector, Security Memory Bridge, Smart Router, Admin Console, Red Team Agents, Ingress Guard, RBAC Middleware, Smart Redaction Engine. Distribution: CLI installer, pre-commit hooks, GitHub Action, admin dashboard, Docker.