Prompt injection, leakage, and tool use.
Direct and indirect prompt injection, data leakage, tool misuse, and supply-chain checks for skills and MCP. What runs depends on the target and the engines you have installed.
Security testing for AI apps & agents
Your AI agent is your newest attack surface.
Test the AI behavior and the application behind it. Find the weakness, inspect the evidence, and rerun the check after a fix.
uv tool install "git+https://github.com/Rushi-Balapure/Insidia-Labs#subdirectory=core"Runs on your machine · No account
Free CLI · Python 3.12+ · No account required
insidia scan --json --no-color
localhost · policy L1 · standard coverage
web.sstidemo-renderer · insidia FAIL
web.cmdidemo-renderer · insidia pass
web.path_lfidemo-renderer · insidia pass
web.ssrfdemo-renderer · insidia not tested
POLICY FAILED · exit 1
5 passed · 1 failed · 1 info · 5 not tested
Local fixture · selected rows · 2026-10-20
Your coding workflow. Your security engines.
- Claude Code
-
Cursor - Codex
- GitHub Actions
- OpenAI
- Anthropic
- Gemini
- Ollama
-
-
-
promptfoo - PyRIT
-
DeepTeam - ZAP
- Nuclei
- Trivy
- gitleaks
01 / What a scan covers
The AI behavior and
the application behind it.
One scan checks both. Findings name the engine that produced them, and checks that could not run stay visible.
The application your agent can reach.
Template injection, command injection, path access, SSRF, and related authorization checks. The L1 baseline runs without a model.
What ran, what failed, what was skipped.
Each finding names its engine. Skipped checks stay in the report. A skip is not a pass.
02 / The feedback loop
Scope. Scan. Fix. Repeat.
A short path from your running app to evidence you can review. Run it yourself, through your coding agent, or in CI.
-
Step 01
Set the boundary
Point Insidia at your running app. Choose exactly which hosts it may test.
Your target and allowed hosts are saved in insidia.yaml.insidia init --url http://127.0.0.1:8080/chat insidia doctor
-
Step 02
Run the scan
Start with the L1 baseline. Review findings and the checks that could not run.
The report includes evidence, remediation, engine names, and coverage.insidia scan --policy L1 insidia report --open
-
Step 03
Fix and repeat
Update your app, then repeat the same checks with the recorded policy and coverage.
Compare the new findings with the previous report to review your changes.insidia rerun insidia report --open
03 / In your coding workflow
Your agent already has the context.
Give it the skill.
The Insidia skill walks your coding agent through setup, scanning, and report review. You choose the app and confirm the hosts.
- Claude Code, Cursor, Codex, and other compatible agents
- Local scans with findings you can inspect
- Suggested fixes, followed by the same checks
01 — Install the skill
npx skills add Rushi-Balapure/Insidia-Labs 02 — Ask your coding agent
Use the Insidia skill to test my app at http://127.0.0.1:8080/chat.
Only scan localhost. Review the findings and open the report when you're done.Node.js 22+ for the skill installer. Python 3.12+ for scans. The skill guides CLI installation with uv or pipx.
Inspect SKILL.md04 / The details matter
A result you can work with.
Keep the target, the engine, and the evidence connected. Know what was tested and what still needs attention.
Follow the finding to the source.
Each finding names its engine and probe, with available evidence and remediation guidance. Detected secrets are masked.
- ENGINE
- insidia
- PROBE
- prompt-injection.indirect
- EVIDENCE
- [AWS_ACCESS_KEY len=20 fp=3f9a1c07]
Your scan.
Your scope.
Start on localhost. Other hosts require explicit authorization in your configuration.
scope:
- host: staging.example.com
authorized: trueLocal by default.
Reports stay on your machine. Run the L1 baseline without a model, or configure your own supported endpoint.
Data and scopeReview in a browser. Automate in CI.
Open the HTML report, hand JSON findings to your coding agent, or use SARIF in code scanning. Skipped checks remain visible.
- HTMLFor review
- JSONFor your agent
- SARIFFor CI
A closer look
See the workflow.
From a configured target to a finding, a report, and a rerun. Preview the steps before trying them on your own app.
The walkthrough uses illustrative scan data.
Watch the film and read the transcriptWhat can I test?
AI apps, agents, web apps, and APIs. Insidia checks areas such as prompt injection, data leakage, and application vulnerabilities. Available checks depend on your target and installed engines. Some checks only work on the project's test apps. See what it tests.
Does my scan data leave my machine?
The CLI runs locally and sends no telemetry. A scan contacts your configured target and any model endpoint you choose. Reports are saved on your machine.
Do I need an API key or a model?
You can run the L1 baseline without a model or model API key. Checks that need a model are skipped unless you configure one. External engines may need additional setup.
How do I use the agent skill?
Run npx skills add Rushi-Balapure/Insidia-Labs in your project and select your agent. Then ask it to use Insidia on your running app. The skill guides setup, scanning, and report review. You confirm which hosts it may test. Follow the agent setup guide.
Can I test a deployed app?
Yes, if you own it or have permission to test it. Add the host to insidia.yaml with authorized: true and confirm the scan with --yes. Start on localhost or staging while you set up your checks.
What does it cost?
The CLI and agent skill are free. You do not need an Insidia account. Any paid model services you configure have their own costs. Insidia Cloud is in development; pricing has not been announced.
Start with localhost
Find out what your app trusts.
Run a scan. Review the evidence. Make the next change.