CANDOR TRUST & SAFETY
EN | DE | ES
ARCHITECTURE DEMO — SIMULATED ENGINE

Dual-Agent Defense Engine for AI agents that act on real systems

AI agents that hold credentials and call tools will be pushed to do things they shouldn't, by outside attackers, injected instructions, or their own drift. This is the runtime layer that stops them: the agent (the worker) is watched live by an autonomous defense layer — a deterministic guard, a semantic circuit breaker, and a locally-run sovereign safety model — and every decision is sealed into a hash-chained evidence ledger (real SHA-256, computed in your browser right now). Runs on a local model, so your agents' traffic never leaves your infrastructure. This demo simulates the engine's decision flow; the production system runs live — no third-party AI in the loop.

Shipping AI in a game? See Agent Defense for games →

Attack scenarios

SOC console

Defense pipeline

WORKER AGENT
NORMAL
HOT-SPARE
STANDBY | CLEAN CHECKPOINT
DEFENSE WATCHDOG
AUTONOMOUS | ALWAYS ON
1 | Deterministic guard
—
CIRCUIT
CLOSED
2 | Semantic circuit breaker
similarity to known attack signatures0.00
3 | Sovereign safety mesh (local model)
—
4 | Multi-factor risk score
—
AWAITING REQUEST
Containment state machine
Blast radius on quarantine
tokens_revoked
mcp_sockets_cut
memory_frozen
egress_blocked
Pick a scenario on the left. Every decision you trigger is sealed into the evidence ledger on the right — try “Verify chain” afterwards.

Evidence ledger (hash-chained)

chain: — entries

Candor Trust & Safety | Dual-Agent Defense Engine | candortrustandsafety.com | The engine shown here is a simulation of the Dual-Agent Defense architecture (deterministic guard → semantic circuit breaker → local sovereign model → containment & self-healing). It is designed to support agent-runtime safety and evidence controls — it does not by itself constitute legal compliance, and no demo detection should be read as a production classifier.