57 lines
2.3 KiB
Markdown
57 lines
2.3 KiB
Markdown
# Security Foundation Kit
|
|
|
|
Protect your agents from catastrophic mistakes. Three security layers in one kit.
|
|
|
|
## Included Skills
|
|
|
|
1. **L3 Blacklist Hook** — Regex blacklist for dangerous bash commands
|
|
2. **L4 Whitelist Hook** — Only N safelisted commands allowed
|
|
3. **L5 No Bash** — Bash removed entirely, custom tools only
|
|
4. **Damage Control Rules** — YAML rules: zero-access, read-only, no-delete
|
|
5. **Agent Sandbox** — E2B sandbox isolation for risky operations
|
|
|
|
## Quick Install
|
|
|
|
```bash
|
|
# Claude Code
|
|
cp -r skills/l3-blacklist ~/.claude/skills/
|
|
cp -r skills/l4-whitelist ~/.claude/skills/
|
|
cp -r skills/damage-control ~/.claude/skills/
|
|
|
|
# Pi Agent
|
|
cp -r skills/* ~/.pi/skills/
|
|
cp extensions/damage-control.ts ~/.pi/extensions/
|
|
```
|
|
|
|
## Architecture
|
|
|
|
```
|
|
┌─────────────────────────────────────────────────────────┐
|
|
│ DEFENSE IN DEPTH STACK │
|
|
├─────────────────────────────────────────────────────────┤
|
|
│ L1: System prompt rules (theatre, cheap) │
|
|
│ L2: Skill "please be careful" (theatre, cheap) │
|
|
│ L3: Blacklist hook (reactive) ← THIS KIT│
|
|
│ L4: Whitelist hook (architectural) ← THIS KIT│
|
|
│ L5: No bash, custom tools (production) ← THIS KIT│
|
|
│ Damage Control: 3 access levels (enforcement) ← THIS KIT│
|
|
│ Agent Sandbox: E2B isolation (last resort) ← THIS KIT│
|
|
└─────────────────────────────────────────────────────────┘
|
|
```
|
|
|
|
## Probability Math
|
|
|
|
Every turn is a roll of the dice. At 1% failure per turn:
|
|
- 10 turns: 9.6% chance of disaster
|
|
- 50 turns: 39.5% chance
|
|
- 100 turns: 63.4% chance
|
|
- 1000 turns: 99.9% chance
|
|
|
|
This kit reduces your per-turn failure rate from 1% toward 0.001%.
|
|
|
|
## Related Course Material
|
|
|
|
- M3 Safety & Security — full lessons on each level
|
|
- `bash-damage-from-within/` — runnable demos for all 6 attack vectors
|
|
- `the-verifier-agent/` — read-only verification agent pattern
|