dscan-security 0.2.0

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

dscan

An open source agent security suite. Trace and redact your agent's tool calls, scan its prompts and MCP configs, and detect suspicious call chains — then inspect everything in a local dashboard.

pip install dscan-security

The package installs as dscan-security; the import path and CLI stay dscan (import dscan, dscan --help).

Quick start

from dscan import watch

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

Run your agent, then open the dashboard:

dscan dashboard

Tool calls are written as redacted NDJSON to ~/.dscan/traces/ (override with DSCAN_TRACES_DIR). Wrapping any tool with @watch.tool traces its params and result; if you use the Anthropic SDK, @watch also intercepts messages.create() and traces each tool_use block.

What dscan catches

Capability	What it detects
@watch (tracing)	Every tool call — name, params, result, duration — written as redacted NDJSON; flags calls whose params contain secrets
Redaction	AWS keys, Anthropic/OpenAI/GitHub/Stripe tokens, JWTs, emails, phone numbers, SSNs, Luhn-valid credit cards, database-URL passwords, high-entropy secrets
dscan scan	Permissive prompts, injection-prone prompts, hardcoded secrets, excessive tool scope; unverified, overprivileged, and credential-leaking MCP servers
dscan trail	Suspicious tool-call sequences across an agent run (exfiltration, recon, injection relay, data staging, goal drift)
dscan dashboard	Local web UI: sessions, per-call timeline, redacted values, secrets flags, and trail findings

Commands

dscan scan

Static analysis of system prompts and MCP config files in a directory. Exits 1 if any high-severity issue is found.

$ dscan scan ./prompts
                                  HIGH (1)
┏━━━━━━━┳━━━━━━━━┳━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┓
┃ Rule  ┃ File   ┃ Line ┃ Message                     ┃ Snippet           ┃
┡━━━━━━━╇━━━━━━━━╇━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━┩
│ SP003 │ sp.txt │    2 │ Hardcoded secret in prompt  │ ...sk-ant-api03... │
└───────┴────────┴──────┴─────────────────────────────┴───────────────────┘
✗ 1 high-severity finding(s).

dscan watch

@watch is a decorator, not a runtime command. This prints a reminder:

$ dscan watch
⚠ Add @watch to your agent function. See README for usage.

dscan trail

Reads trace files and reports suspicious tool-call chains, grouped by severity. Exits 1 on any high or critical finding. Supports --min-severity {low|medium|high|critical} and --json.

$ dscan trail ~/.dscan/traces/
┏━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ Severity ┃ Pattern         ┃ Tools Involved          ┃ Message       ┃ Confidence ┃
┡━━━━━━━━━━╇━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ CRITICAL │ INJECTION_RELAY │ search_web → send_email │ untrusted...  │        85% │
│ HIGH     │ EXFIL_SEQUENCE  │ read_file → send_email  │ read then...  │        80% │
└──────────┴─────────────────┴─────────────────────────┴───────────────┴────────────┘
2 findings across 5 tool calls analysed

dscan dashboard

Serves the local web UI at localhost:4321. --port sets the port; --no-open skips opening a browser.

$ dscan dashboard --no-open
✓ Dashboard at http://127.0.0.1:4321  (Ctrl-C to stop)

How trail works

dscan trail reads your trace files and looks at the order of tool calls, not just each call on its own. It groups calls by agent run and reports five kinds of suspicious sequences: reading sensitive data and then sending it somewhere external (exfiltration), probing for permissions and then running a tool the agent never declared (reconnaissance), pulling in untrusted web or email content and then immediately sending or executing (injection relay), reading several different sensitive sources in a row with nothing sent in between (data staging), and taking destructive or outbound actions that a read-only goal never asked for (goal drift). Each finding carries a severity and a confidence score so you can triage.

Contributing

git clone https://github.com/DeepScan-Security/dscan
cd dscan
pip install -e ".[dev]"
pytest

Built test-first; coverage stays at or above 80%. See CONTRIBUTING.md for what we need help with, and SECURITY.md to report a vulnerability.

License

MIT