AgnosticSecurity
DLP for AI coding agents. Stops secrets, credentials, and PII from leaking through Copilot, Claude Code, Cursor, and 17+ LLM providers — before the data leaves your machine.
What It Does
-
Makes sensitive data invisible to AI. File gates, context boundaries, and prompt guards ensure your
.envfiles, API keys, and PII never enter an AI context window. If AI never sees it, there is nothing to leak. -
Catches what other tools miss. 19 detection layers handle base64-encoded exfiltration, split-across-turns attacks, tool call smuggling, memory poisoning, and cross-session taint tracking. 242 detection patterns across 64 attack categories.
-
Works with every AI tool you already use. One install protects Copilot, Claude Code, Cursor, Windsurf, ChatGPT, and LangChain agents simultaneously. No vendor lock-in, no code changes.
Quick Start
pip install securityagent-core
as-init # auto-detects AI tools, configures protections — 30 seconds
That's it. Your .env files, credentials, and PII are now protected across every AI tool in your environment.
Other install methods: Docker | VS Code Extension | Helm | GitHub Action
How It Works
Developer using Copilot / Claude Code / Cursor / ChatGPT / LangChain
|
| AI tries to read .env ──────────── BLOCKED (file gate, <1ms)
| Prompt contains SSN ────────────── BLOCKED (prompt guard, 10 decoding layers)
| Agent runs `curl -d @.env` ─────── BLOCKED (exec guard + egress allowlist)
| @workspace indexes credentials ─── BLOCKED (context boundary)
| LLM response leaks PII ────────── BLOCKED (output pipeline)
| Tool call smuggles data out ────── BLOCKED (tool call guard + taint tracker)
| Memory stores "password is X" ──── BLOCKED (memory guard)
| Base64-encoded exfil in 3 steps ── BLOCKED (multi-step exfil detector)
| External agent probes your API ─── BLOCKED (6-layer ingress guard)
|
└── Everything else ──────────────── Works normally
Key Numbers
| Metric | Value |
|---|---|
| Detection patterns | 242 across 19 layers (tiered T0/T1/T2) |
| Attack categories | 64/64 covered (100%) |
| HarmBench ASR | sub-3%, 0% false positives, F1 0.98+ |
| Red team agents | 55 Docker-isolated agents, 84 events, 100% detection |
| Compliance | 65/65 — OWASP 10/10, NIST 26/26, MITRE 14/14, CSA 15/15 |
| Tests | 1,235 across 32 suites (100% pass) |
| Languages | 40 (multilingual prompt injection detection) |
| LLM providers | 17+ supported (OpenAI, Anthropic, Google, Azure, Ollama, ...) |
| Components | 48+ |
Features
Detection & Prevention
- File gate — blocks AI from reading
.env,.pem,id_rsa, credentials - Prompt guard — 10 decoding sublayers (base64, hex, ROT13, unicode, zero-width chars, reversed text, leetspeak, digit reassembly)
- Exec guard — catches
cat ~/.env,curl -d @secrets, obfuscated payloads, netcat/SSH tunnels - Context boundary —
.copilotignore+search.exclude+ LM API interceptor ensures AI never sees sensitive content - Terminal Guard v4 — process ancestry tree + input cadence detection separates AI from human terminal commands
- Output pipeline — scans LLM responses for leaked PII, credentials, semantic disclosure patterns
- Taint tracking — SHA-256 + n-gram Jaccard similarity detects exfiltration across sessions
- Tool call guard — DLP + taint scan on MCP/function call arguments
- Lethal trifecta detector — blocks when private data + untrusted input + external comm all active simultaneously
- Streaming DLP — real-time scanning of streamed LLM responses
- Memory scanner — detects trust boundary violations and memory poisoning
- Injection detection — 242 regex patterns covering SQL injection, SSRF, SSTI, XXE, Log4Shell, deserialization, and more
- Code fingerprint guard — proprietary code leak prevention via n-gram fingerprinting
- Vulnerability scanner — OWASP SAST-lite for AI-generated code
Governance & Compliance
- RBAC — 7 roles (super_admin through developer), team scoping, API key resolution
- Smart redaction — 4 modes per data type: block, redact (reversible tokenization), mask (partial), allow
- Privacy mode —
full_privacy(local LLM only),balanced,permissive - Policy-as-code — YAML policy templates, versioned, org-wide + per-team overrides
- Alert rules — configurable thresholds and notification channels
- Compliance scoring — OWASP, NIST AI RMF + 800-53, MITRE ATLAS, CSA ARIA
- Compliance PDF export — enterprise-styled reports with score cards, per-framework detail, eval metrics
- Audit backend — PostgreSQL (production) or SQLite (local), immutable SHA-256 checksummed logs
- Security scorecard — real-time posture assessment across all detection layers
Dashboards (7 pages, unified design system)
- Dashboard — gateway stats, breach data, DLP blocks, latency metrics
- Prompt Injection — real-time injection attempt visualization
- Red Team — attack replay + forensics from Docker harness
- Events — paginated audit trail with filters (action, model, search)
- DLP Controls — per-type PII blocking, egress domain allowlist, redaction policy
- Console — policy management, agent enrollment, RBAC administration
- Platform — SDK status, cost tracking, system health
Integrations
- CrowdStrike EDR — IOC push on high-severity blocks
- Splunk HEC — audit event forwarding
- Datadog — metrics and event ingestion
- PagerDuty / OpsGenie — alert escalation
- Slack / Teams — real-time alert bot
- WebSocket event feed — live streaming for custom dashboards
CISO Skills (Claude Code slash commands)
8 self-contained security workflows — no code deps, call existing APIs:
| Command | What It Does |
|---|---|
/ciso |
Executive security briefing with risk assessment |
/threat-brief |
Current threat landscape analysis |
/audit-now |
Run 50-check security audit |
/redteam-auto |
Automated red team exercise |
/posture |
Security posture scorecard |
/incident |
Incident response workflow |
/learn-security |
Interactive security training |
/compliance-check |
Multi-framework compliance validation |
Supported AI Tools
| AI Tool | Protection Method |
|---|---|
| GitHub Copilot | VS Code extension (context boundary + LM API interceptor) |
| Claude Code | Hook-based (PreToolUse + UserPromptSubmit, 4-layer pipeline) |
| Cursor | .cursorignore + VS Code extension |
| Windsurf | .aiignore + VS Code extension |
| ChatGPT, Claude.ai, Gemini, Grok, Perplexity, DeepSeek | Chrome extension (DLP for web LLM interfaces) |
| LangChain / Autogen | Auto-instrumentation SDK (zero-code) |
| Any LLM provider | API Gateway with input/output security pipelines |
Installation
| Method | Command | Best For |
|---|---|---|
| PyPI | pip install securityagent-core && as-init |
Individual developers |
| Docker | docker compose up |
Full stack (gateway + breach engine + proxy) |
| VS Code | Install .vsix from Releases |
Copilot / Cursor users |
| Chrome | Load unpacked from chrome-extension/ |
ChatGPT / Claude web users |
| Pre-commit | bash hooks/install.sh |
Git-level scanning |
| GitHub Action | See below | CI/CD pipeline |
| Helm | helm install agsec helm/agnosticsecurity/ |
Kubernetes production |
GitHub Action
name: AgnosticSecurity Scan
on: [pull_request]
jobs:
scan:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: secure-mind-live/AgnosticSecurity/.github/actions/security-scan@main
with:
severity-threshold: HIGH
fail-on-findings: true
scan-dlp: true
scan-vulns: true
Architecture
+------------------+
| Developer IDE |
| (VS Code/Cursor) |
+--------+---------+
|
+-------------------+-------------------+
| | |
+--------v--------+ +-------v--------+ +--------v--------+
| VS Code Ext | | Claude Code | | Chrome Ext |
| Context boundary| | Hooks (4-layer)| | Web LLM DLP |
+---------+--------+ +-------+--------+ +--------+--------+
| | |
+-------------------+-------------------+
|
+--------v---------+
| API Gateway | Port 8000
| (Smart Router) |
+--------+---------+
|
+-------------------+-------------------+
| | |
+--------v--------+ +-------v--------+ +--------v--------+
| Input Pipeline | | Output Pipeline| | Ingress Guard |
| DLP + injection | | PII + redaction| | 6-layer defense |
| + guardrails | | + de-tokenize | | + risk scoring |
+--------+--------+ +-------+--------+ +--------+--------+
| | |
+-------------------+-------------------+
|
+-------------+-------------+
| |
+--------v--------+ +--------v--------+
| Breach Engine | | Audit Backend |
| Classification | | PostgreSQL/ |
| + Compliance | | SQLite |
+-----------------+ +-----------------+
Testing
source .venv/bin/activate
# Quick validation
python3 scripts/smoke_test.py # 31 tests — file gate + content DLP
python3 scripts/test_red_team.py # 52 tests — adversarial attacks
# Red team Docker harness (55 agents, 10 OWASP categories)
docker compose -f docker-compose.redteam.yml up --build -d
docker exec agsec-attacker python3 /agents/run_all.py
# Full suite: 1,235 tests across 32 files — see CLAUDE.md for complete list
Privacy
- Runs on your machine — no telemetry, no analytics, no cloud dependency
- Pluggable LLM — local Ollama (air-gapped), Anthropic, or OpenAI for optional semantic analysis
- Privacy mode —
full_privacylocks to local models only;balancedallows cloud with DLP;permissivelogs only - Audit logs stored locally — SHA-256 checksummed, tamper-detected
Documentation
| Document | Description |
|---|---|
| Security Design | Threat model, design decisions, failure modes |
| OWASP LLM Mapping | Control-to-risk compliance mapping |
| Admin Console | Policy management architecture |
| Enterprise RBAC | Role-based access control design |
| Memory Agent Threats | Memory poisoning, trust boundaries |
| Competitive Analysis | vs Lakera, Protect AI, Prompt Security, HiddenLayer |
Contributing
git clone https://github.com/secure-mind-live/AgnosticSecurity.git
cd AgnosticSecurity
python3 -m venv .venv && source .venv/bin/activate
pip install -e .
python3 scripts/smoke_test.py # verify setup
PRs welcome. All contributions run through the security hook — if it blocks your action, adjust your approach.
License
Business Source License 1.1 (BSL-1.1). Free for non-production use. See LICENSE for details.
Metadata
Release files for agnostic-security 4.61.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| agnostic_security-4.61.0.tar.gz | 245.7 kB | Without provenance | Details for agnostic_security-4.61.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| agnostic_security-4.61.0-py3-none-any.whl | Python 3 | none | any | Details for agnostic_security-4.61.0-py3-none-any.whl |
Total release size: 489.8 kB
Release files / agnostic_security-4.61.0.tar.gz
| Download URL | agnostic_security-4.61.0.tar.gz |
|---|---|
| Size | 245.7 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
e637f499bdcabb805a569de4b6125d68eb2cc5624cdf21e91a07c9ddd10b4ea3
|
|
BLAKE2b-256 checksum How to use checksums |
d65b94a2ffd26d2a5c034e229dfe79797411671cb68e0676535ed78a83ec3824
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.12.11
|
Release files / agnostic_security-4.61.0-py3-none-any.whl
| Download URL | agnostic_security-4.61.0-py3-none-any.whl |
|---|---|
| Size | 244.1 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
af4d58eef9311f15d8a0a0d48bf0b4ea3acd7f993ed725da9b581ac9d2a3f3ca
|
|
BLAKE2b-256 checksum How to use checksums |
b23829295eab66df725c2d42633bf047887de8f312f7a80fd04a7887252834c0
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.12.11
|
Security for agnostic-security 4.61.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.