Open-source · Zero-dependency · OWASP-anchored

The calm guardian for
AI, agent & MCP code.

A standards-anchored AI security toolkit any developer or reviewer can run in seconds — in the IDE, in CI, or from the terminal. You never rewrite your LLM call. You wrap it.

Deterministic by design — no LLM in the loop. Same code, same verdict, every run; fully offline, private, and free. Your gate never flakes and your source never leaves your machine.

$ pip install grey-panda

★ Star on GitHub Read the docs

Apache-2.0 Python 3.9+ runtime deps 0 deterministic · no LLM 26 scanner rules LLM · Agentic · AISVS · MCP

The problem

AI ships a new attack surface — mostly unguarded.

Traditional AppSec never covered prompts, tools, agents, or MCP. These are the risks your existing scanners don't see.

LLM01

Prompt injection is #1

The top AI attack pattern — and it needs no authentication. Any user input can hijack the model.

LLM03 · ASI

Agents act irreversibly

One injected instruction can make an agent send money, delete data, or call a tool with your credentials.

AISVS C10

MCP tool poisoning

A tool's description hides instructions, or a trusted tool is silently swapped after you approved it — a "rug pull".

The bundle

Five kits. One zero-dependency engine.

Each audience gets a clear front door — all powered by one shared, tested engine. No duplication, no drift.

→

Module 1

🧰

Developer Kit

Drop-in SDK guardrails + IDE integration. Protect an LLM call in under 2 minutes.

→

Module 2

🛡️

Security Reviewer Kit

AppSec gate: AISVS verify, threat models, injection battery, sign-off.

→

Module 3

🔍

Scanner & CI/CD Kit

26 rules → OWASP IDs. Markdown / JSON / SARIF. GitHub Action + pre-commit.

→

Module 4

🤖

MCP & Agent Security Kit

Rule of Two, HITL, kill switch, MCP manifest pinning, ACS Guardian.

→

Module 5

📚

Standards & Governance Kit

The knowledge pack, mappings, and the honest Can/Cannot-Do doc.

One CLI drives them all —
scan · verify · mcp · init

How it works

A 7-step request pipeline.

Defense in depth: known patterns blocked at input, data quarantined even if injection succeeds, output cleaned, every action audited.

STEP 1

PromptGuardrail

block injection

STEP 2

DLPScanner

redact PII & secrets

STEP 3

SecureContext

tag trust

STEP 4

LLM call

your call — unchanged

STEP 5

DLPScanner

scan the response

STEP 6

OutputGuardrail

escape → XSS-safe

STEP 7

AuditLogger

zero-raw-text event

Building an agent?

Add AgentSecurityWrapper — the Rule of Two enforced at construction, deny-by-default tools, HITL gates, per-session budgets, and a kill switch. See the architecture →

Credibility

Standards-anchored, not opinion-driven.

Every rule, checklist item, and SDK control cites a specific ID. Also grounded in NIST AI 100-2 and Meta's "Rule of Two".

OWASP LLM Top 10 2026

LLM01–LLM10

GenAI Data Security

DSGAI01–DSGAI21

Agentic Apps Top 10

ASI01–ASI10

AISVS

C1–C12 · L1/L2/L3

MCP Security Guides

dev + third-party

Agent Control Standard

hooks · dispositions · AgBOM

Proof

Spotless by construction.

Grey Panda scans its own repository clean in CI. It practices exactly what it preaches.

scanner rules

runtime dependencies

tests, all green

formats: md·json·sarif

20 findings → 0 ✓ the same app, rebuilt with Grey Panda controls

For everyone

Same safety floor. Scaled process.

One tool, three profiles — from a solo indie dev to a regulated enterprise. Nobody is priced out of safety.

solo

Indie

Prototypes & side projects

High-signal core rules
Drop-in SDK
Fail on CRITICAL

team

Startup

Teams shipping fast

+ DLP, RAG, MCP, shadow-AI
+ CI gate & pre-commit
Fail on HIGH

enterprise

Regulated

Large / compliance-bound

Every rule on
+ AppSec gate & AISVS L1/L2/L3
Fail on HIGH

🧭 Honest about limits

Grey Panda is a strong floor, not a ceiling. Pattern matching can't stop all prompt injection; regex DLP is language-specific; static analysis has false positives and negatives. We ship a whole document — with a confidence level and failure condition for every capability. Read what it can & cannot do →

Get started

Two minutes to safer AI.

$ pip install grey-panda && gp scan .

★ Star on GitHub View on PyPI Read the docs