securityagent-core
The DLP engine behind AgnosticSecurity -- the firewall for AI coding agents.
v4.60.0 | 1,235 tests | 242 detection patterns | 19 layers | 64/64 attack categories | 65/65 compliance | 40 languages
Quick Start
pip install securityagent-core
as-init
as-init auto-detects your AI tools and configures protections. Supports:
Claude Code | Copilot | Cursor | Windsurf | Aider | Continue | Tabnine | Codeium
What's Inside
| Package | What It Does |
|---|---|
endpoint_agent/ |
DLP engine -- file blocking (60+ path patterns), confidence-scored content scanning (PII, credentials, semantic disclosure) with structural validators (Luhn, SSA, entropy), data flow taint tracking (hash/n-gram/substring), tool call argument scanning (MCP/function DLP), behavioral monitoring, honeypot traps. Notebook-aware .ipynb + PDF extraction (PyMuPDF/Tesseract/pypdf). Image OCR via Tesseract + QR/barcode decode. |
skills/ |
MCP server (9 tools) -- secure_read, secure_write, secure_exec, analyze_prompt, scan_output, check_policy, get_session_policy, audit_log, scan_code. Works with any MCP-compatible client. |
policy/ |
Policy engine -- per-session least privilege, behavioral chain detection (11 attack patterns), audit trail with agent attribution. Obsidian vault integration for cross-session policy context. |
plugin.py |
CLI entry point -- validate_exec(), validate_prompt(), validate_output() pure functions. |
Key Features
Exec Guard (60+ rules)
Blocks dangerous commands before execution. High-confidence rules (env var dumps, encoded exfil) block immediately. Low-confidence rules go through a 4-tier decision: domain allowlist, localhost check, payload scan, LLM verification.
Prompt Scanner (10 decode layers)
Catches PII and credentials even when obfuscated: direct regex, base64, hex escapes, ROT13, unicode escapes, URL percent-encoding, zero-width character stripping, reversed text, word-to-number with leetspeak, partial credential completion. Plus digit fragment reassembly and low-sensitivity PII escalation.
Terminal Guard
Traces process ancestry -- if a parent process is an AI agent and a child process touches sensitive files, the operation is blocked. Priority-based cadence scanning.
Code Fingerprint Guard
Registers proprietary code via n-gram fingerprinting. Detects similar code in prompts even with renamed variables (Jaccard similarity). Locality-aware thresholds.
Data Flow Taint Tracking
Tags sensitive data at ingress (file read, prompt, tool result). Detects at egress via SHA-256 hash, n-gram Jaccard similarity, and substring matching. Cross-session persistence (24h TTL).
Install
# From PyPI
pip install securityagent-core
# From source
pip install git+https://github.com/secure-mind-live/securityagent-core.git
# With optional dependencies
pip install "securityagent-core[cloud]" # Cloud LLM providers (Anthropic, OpenAI)
pip install "securityagent-core[llm]" # Local LLM (Ollama)
pip install "securityagent-core[dev]" # Test dependencies
Usage
Python SDK
from skills.adapters.python_sdk import SecurityAgentSDK
sdk = SecurityAgentSDK(agent_id="my-agent", agent_type="langchain")
result = sdk.secure_read("/path/to/file") # DLP-gated file read
result = sdk.secure_exec("ls -la") # Command validation
result = sdk.analyze_prompt("Get all SSNs") # Prompt intent analysis
MCP Server
python -m skills.adapters.mcp_server
Any MCP-compatible client (Claude Code, Copilot, Cursor, custom agents) connects via stdio JSON-RPC.
CLI
python -m plugin ~/.env # File read gate
python -m plugin --exec "cat ~/.env" # Exec validation
python -m plugin --prompt "Get all passwords" # Prompt analysis
python -m plugin --mcp-server # MCP server mode
Claude Code Integration
Add to .claude/settings.json:
{
"hooks": {
"PreToolUse": [
{ "command": "python3 /path/to/hook_fast.py" }
]
}
}
Or run as-init -- it configures this automatically.
Tests
pip install -e ".[dev]"
pytest # Unit tests
python3 scripts/smoke_test.py # 31 integration tests
Full suite: 1,235 tests across 32 suites -- see AgnosticSecurity for the complete test runner.
Numbers
| Metric | Value |
|---|---|
| Detection patterns | 242 |
| Detection layers | 19 (tiered: T0/T1/T2) |
| Attack categories | 64/64 (100%) |
| Compliance controls | 65/65 -- OWASP, NIST, MITRE, CSA |
| HarmBench ASR | Sub-3% |
| False positive rate | 0.99% |
| Languages | 40 |
License
Proprietary. See LICENSE.
Metadata
Release files for securityagent-core 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 |
|---|---|---|---|---|
| securityagent_core-4.61.0.tar.gz | 251.8 kB | Without provenance | Details for securityagent_core-4.61.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| securityagent_core-4.61.0-py3-none-any.whl | Python 3 | none | any | Details for securityagent_core-4.61.0-py3-none-any.whl |
Total release size: 533.2 kB
Release files / securityagent_core-4.61.0.tar.gz
| Download URL | securityagent_core-4.61.0.tar.gz |
|---|---|
| Size | 251.8 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
0966a0eeb2790b4ea779d0b848ef20ac741ae042ab2a2f37711851140b52b6eb
|
|
BLAKE2b-256 checksum How to use checksums |
2f361e2b8ce54dcd0b297f607cf0d9e9e728a2747393e94909cc1d21425fab89
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.12.11
|
Release files / securityagent_core-4.61.0-py3-none-any.whl
| Download URL | securityagent_core-4.61.0-py3-none-any.whl |
|---|---|
| Size | 281.4 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
30cec809816cae24456fa62b582c9c6d14b89aa493395ea2273276565a83a67c
|
|
BLAKE2b-256 checksum How to use checksums |
cf84ffa03307891c30bda1231b2f1945e224ef2cc2cf2b49dece71c764674a95
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.12.11
|
Security for securityagent-core 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.