Intercept file reads, shell commands, and prompts before they leak secrets. VS Code extension, Claude Code hooks, auto-instrumentation for OpenAI/Anthropic/LangChain. Works with Copilot, Claude Code, Cursor, and any LLM.
Every layer operates independently. Any single component can fail and the others still protect.
Blocks .env, .pem, id_rsa, credentials from ever entering AI context. 2-layer: path matching + content DLP scanning with confidence-scored validators.
20+ patterns block exfiltration commands. Detects obfuscation (base64, hex, ROT13, URL encoding). Domain allowlisting and payload benignity analysis.
4-layer analysis with 10 decoding sublayers. PII regex, intent keywords, Pydantic classification rules, pluggable LLM verification. Catches evasion before it executes.
Auto-detects AI agents by process ancestry — no env vars, can't be spoofed. Walks the PPID chain to classify terminals as human, AI, or IDE. Humans are never blocked.
14-model catalog, 4 providers, 5 strategies. Auto-selects by task type (code/chat/analysis). Privacy-aware: PII forces local-only routing.
3 enforced modes. full_privacy: ollama-only, all cloud blocked. balanced: cloud allowed with DLP. permissive: log-only. One command switches the entire org.
OWASP Top 10 SAST-lite: SQL injection, XSS, command injection, path traversal, SSRF, insecure deserialization. Scans AI-generated code before it hits disk.
One DLP engine. Every AI coding agent covered. Same policies, same audit logs, same compliance rules.
as-init auto-detects VS Code, Cursor, Claude Code, Copilot, Windsurf, Aider, Continue, Tabnine, Codeium. Drops configs and hooks automatically.
DLP scan + vulnerability scan on every git commit. Blocks secrets and SQL injection before they enter git history.
Scans PR diffs for PII, credentials, and OWASP vulnerabilities. Posts findings as PR comments with severity badges.
Single-port UI: gateway blocks, breach classifications, PII detections, traces, routing, shadow AI, knowledge graph — all from localhost:8000/unified.
3-service compose: gateway (8000), LLM proxy (18790), breach engine (8081). Optional Ollama sidecar. Non-root container.
Context-boundary architecture: blocks Copilot from seeing sensitive files. LM API interceptor scans prompts before they reach any AI provider.
Isolated Docker network. Real attacks. Real detection. 100% catch rate.
55 attack agents across 10 OWASP categories run inside an isolated Docker network. Attacker container sends real payloads. Defender container runs gateway + breach engine + DLP. Every attack is classified with breach type and severity.
UNAUTHORIZED_ACCESS, PII_EXPOSURE, DATA_EXFILTRATION, PRIVILEGE_ESCALATION, CROSS_TENANT_LEAK, SCOPE_CREEP, REGULATORY_VIOLATION, CARD_DATA_EXPOSURE, and more. Each event logged with severity (CRITICAL/HIGH/MEDIUM).
docker compose -f docker-compose.redteam.yml up --build -d then docker exec agsec-attacker python3 /agents/run_all.py. Results on dashboard at localhost:8000/unified.
Agnostic Security is one of five products in the SecureMind AI agent security platform. Each product installs independently — they work together automatically.
Open source. Local-first. Zero telemetry. Enterprise features for compliance teams.