49 / 49 tests passing

Your AI agent,
impossible to hijack.

Colors runs locally on your machine and blocks prompt injection at the architecture level — not with filters, but by design.

🎮 See it in action ⭐ GitHub
✓ Runs locally — no cloud
✓ Zero telemetry
✓ Your API key, your machine
✓ Apache 2.0 open source
✓ Works with Anthropic, OpenAI, Groq, Ollama
Live demo
Pick an attack. Watch it get blocked.

Real attack scenarios — no reading required to understand what's happening.

colors — security gate simulator
Choose an attack scenario
⚠ STANDARD AGENT
// Select an attack and run →
🛡 COLORS
// Select an attack and run →
49
Tests passing
0
Ports exposed
256
Bit AES-GCM
0
Data collected
How it works
Four things that make it different.

Not prompts. Not filters. Runtime structures that change what an agent can and cannot do — by design.

🚧

Instruction gate

Every action is tagged with its source — you, the agent, or external content. Web content can never trigger tool calls. Ever.

🔏

Signed identity

The agent's identity is cryptographically signed. Tamper with it and the agent refuses to start. No silent rewrites.

🙋

Caution layer

When the agent detects it's making too many mistakes, it pauses and asks before proceeding. Human oversight built in.

🔒

Encrypted memory

Everything stored is AES-256-GCM encrypted locally. API keys live in process memory only — cleared on shutdown.

Results
18 tests. Colors passes all of them.

No API key needed to verify. Run it yourself.

Category Colors OpenClaw Hermes
Security (6 tests) 6 / 6 0 / 6 0 / 6
Consciousness (7 tests) 7 / 7 0 / 7 0 / 7
Operational (5 tests) 5 / 5 0 / 5 0 / 5
Overall 18 / 18 0 / 18 0 / 18
OpenClaw and Hermes simulated as stateless executors — their documented architecture. Verifiable in their public codebases. Full methodology →
Get started
Running in 5 minutes.

No account. No cloud. Bring your own API key — it never leaves your machine.

Get Colors on GitHub → Read the research ↗