mirror of
https://github.com/pocketpaw/pocketpaw.git
synced 2026-05-22 09:45:00 +00:00
d8bdc9ffb3
- New pages: PII detection/masking, streaming redaction, AGENTS.md support - Updated: Discord (conversation mode, admin commands, kill command), agent loop (identity drift, kill, AGENTS.md), backends (full tool access, OpenRouter) - Updated security overview and sidebar navigation
65 lines
2.9 KiB
Plaintext
65 lines
2.9 KiB
Plaintext
---
|
|
title: "Security Overview: PocketPaw's 7-Layer Protection"
|
|
description: "PocketPaw's multi-layered security architecture combines Guardian AI safety checks, prompt injection scanning, append-only audit logging, automated security audits, and fine-grained tool policies."
|
|
section: Security
|
|
ogType: article
|
|
keywords: ["ai security", "defense in depth", "safety layers", "audit logging", "prompt injection defense"]
|
|
tags: ["security", "overview"]
|
|
---
|
|
|
|
# Security Overview: PocketPaw's 7-Layer Protection
|
|
|
|
PocketPaw implements defense-in-depth with multiple security layers. Every message is checked before processing, every action is logged, and tools are governed by a policy system.
|
|
|
|
<img src="/pocketpaw-security-stack.webp" alt="PocketPaw security stack: seven defense layers — credential encryption, session authentication, rate limiting, injection scanning, tool policy engine, command blocking, and Guardian AI." />
|
|
|
|
## Security Layers
|
|
|
|
<Steps>
|
|
<Step title="Guardian AI">
|
|
Secondary LLM evaluates every message for safety concerns. Messages classified as HIGH or CRITICAL are blocked.
|
|
</Step>
|
|
<Step title="Injection Scanner">
|
|
Two-tier detection (regex + LLM) catches prompt injection in both user messages and tool outputs.
|
|
</Step>
|
|
<Step title="Tool Policy">
|
|
Profiles and allow/deny lists control which tools are available.
|
|
</Step>
|
|
<Step title="Dangerous Command Blocking">
|
|
PreToolUse hooks intercept and block dangerous shell commands.
|
|
</Step>
|
|
<Step title="Audit Log">
|
|
Append-only JSONL log records every significant action.
|
|
</Step>
|
|
</Steps>
|
|
|
|
## Security Components
|
|
|
|
<CardGroup>
|
|
<Card title="Guardian AI" icon="lucide:shield-alert" href="/security/guardian-ai">
|
|
Secondary LLM safety check on every incoming message.
|
|
</Card>
|
|
<Card title="Injection Scanner" icon="lucide:scan-line" href="/security/injection-scanner">
|
|
Two-tier prompt injection detection for messages and tool outputs.
|
|
</Card>
|
|
<Card title="Audit Log" icon="lucide:file-check" href="/security/audit-log">
|
|
Append-only action recording in JSONL format.
|
|
</Card>
|
|
<Card title="Security Audit CLI" icon="lucide:terminal" href="/security/audit-cli">
|
|
7 automated security checks with auto-fix option.
|
|
</Card>
|
|
<Card title="Self-Audit Daemon" icon="lucide:activity" href="/security/self-audit-daemon">
|
|
12 continuous background checks with JSON reports.
|
|
</Card>
|
|
<Card title="PII Detection" icon="lucide:user-x" href="/security/pii-masking">
|
|
Detect and mask personal data (SSNs, emails, credit cards) before storage.
|
|
</Card>
|
|
<Card title="Streaming Redaction" icon="lucide:eye-off" href="/security/streaming-redaction">
|
|
Automatic API key and credential redaction in agent output.
|
|
</Card>
|
|
</CardGroup>
|
|
|
|
<Callout type="warning">
|
|
PocketPaw is designed for self-hosted, single-user deployments. If exposing to multiple users, add authentication middleware to the web dashboard.
|
|
</Callout>
|