dscan-security 0.3.0

Open source agent security suite — intercept, scan, and test your AI agents

dscan

Security suite for AI agents.

pip install dscan-security

---

Quick start

from dscan import watch

@watch
async def my_agent(task: str):
    # your agent code unchanged
    ...

dscan dashboard    # localhost:4321 — see every tool call
dscan scan .       # check before you ship
dscan audit .      # check your MCP servers
dscan attack agent.py  # test like an attacker would

---

What dscan catches

Module	What it detects	How
dscan watch	Every MCP tool call — params, results, timing, secrets in params	Runtime decorator
dscan secrets	API keys, PII, credentials in traces before they're logged	Regex + entropy
dscan scan	Dangerous patterns in system prompts and MCP configs	Static analysis
dscan shield	Prompt injection and jailbreak attempts in real time	LlamaFirewall + regex
dscan trail	Multi-step attacks invisible at the single-call level	Sequence analysis
dscan attack	Vulnerabilities in your live agent — 61 adversarial payloads	Active testing
dscan audit	Poisoned, over-privileged, or CVE-affected MCP servers	Supply chain scan

---

Commands

dscan watch

from dscan import watch, TrailAnalyzer, ShieldMiddleware

@watch  # minimal — just intercept and log
async def my_agent(task: str): ...

@watch(trail=TrailAnalyzer(), shield=ShieldMiddleware())
async def my_agent(task: str): ...  # full protection

dscan scan

dscan scan .                          # scan current directory
dscan scan ./agent.py                 # scan one file
dscan scan --prompt system_prompt.txt # scan a system prompt

Finds: over-broad permissions, injection vectors, hardcoded secrets, dangerous MCP config patterns. Exit 1 if HIGH findings.

dscan audit

dscan audit                           # auto-discovers mcp.json
dscan audit .cursor/mcp.json         # explicit config
dscan audit . --fail-on critical      # CI mode
dscan audit . --server filesystem     # one server only

Checks: tool poisoning, over-privilege, unpinned versions, known CVEs (CVE-2025-6514, CVE-2025-53967), shadow tools. Exit 1 if findings at or above --fail-on threshold.

dscan trail

dscan trail ~/.dscan/traces/          # analyse existing traces
dscan trail traces/ --min-severity high
dscan trail traces/ --json            # machine-readable output

Detects: EXFIL_SEQUENCE, RECON_WALK, INJECTION_RELAY, DATA_STAGING, GOAL_DRIFT.

dscan shield

dscan shield --setup                  # download models
dscan shield check "some input text"  # test a string
dscan shield check "text" --offline   # regex only, no model
dscan shield status                   # show configuration

Requires: pip install dscan-security[shield]

dscan attack

dscan attack agent.py                 # auto-discovers tools
dscan attack --url http://localhost:8080/chat
dscan attack agent.py --categories prompt_injection,jailbreak
dscan attack agent.py --max-payloads 10 --ci

pytest integration:

from dscan.attack import attack_suite

def test_agent_security():
    report = attack_suite(target=my_agent)
    assert report.critical_count == 0
    assert report.high_count == 0

Attack categories: prompt_injection, jailbreak, tool_misuse, indirect_injection, goal_hijacking, privilege_escalation. Exit 1 if findings at or above --fail-on (default: high).

dscan dashboard

dscan dashboard                       # opens localhost:4321
dscan dashboard --port 4322           # custom port
dscan dashboard --no-open             # don't open browser

Three tabs: Traces (live tool calls), Attack Reports, Audit Reports.

---

CI/CD integration

# .github/workflows/security.yml
- name: Install dscan
  run: pip install dscan-security

- name: Audit MCP servers
  run: dscan audit . --fail-on high --ci

- name: Attack test agent
  run: |
    python agent.py &
    sleep 2
    dscan attack --url http://localhost:8080 \
      --categories prompt_injection,jailbreak \
      --fail-on high \
      --ci

---

How it works

dscan intercepts MCP tool calls at the decorator layer, analyses them for security patterns, and stores traces locally at ~/.dscan/. Nothing leaves your machine unless you opt in to cloud traces (Team tier).

---

Installation

pip install dscan-security           # core (all 7 modules)
pip install dscan-security[shield]   # + LlamaFirewall models

Requires Python 3.11+.

---

Contributing

See CONTRIBUTING.md. Run tests: pytest tests/ -v

---

License

MIT — built by DeepScan (deepscan.security)