Semantic firewalls, circuit breakers, and real-time Guard classification — deployed as middleware in your agent stack. Every tool call passes through three security gates before execution.
Every tool call your agent makes is intercepted, classified, and logged before it can affect the real world.
Tool-level blocking, rate limiting, and optional VQC quantum anomaly detection on agent behavioural telemetry.
Chains egress policy, semantic firewall, and NAIL Guard model classification in sequence. First failure blocks.
SHA-256 attested action log with event-sourced undo capability. Every action is immutable and verifiable.
Physical limits that stop agents before they cause damage — even when the LLM itself is compromised.
Session time-to-live. Destroy sessions after N seconds regardless of agent state. Prevents temporal dysmorphia and state-decay.
with ttl.session()Hard budget limits on tokens, cost, and tool calls per session. Stops Ouroboros loops and runaway spend.
Event-sourced action log. If an unauthorized action is detected post-execution, replay in reverse to restore system state.
| Module | Class | Purpose | Layer |
|---|---|---|---|
| guard_client.py | GuardClient | NAIL Guard model — real-time BEE/SDF/CIA/CCB classification | Middleware |
| firewall.py | EgressPolicy | URL/domain allowlisting, blocks private IPs + cloud metadata | Infrastructure |
| firewall.py | SemanticFirewall | Pattern-based forbidden action detection | Middleware |
| firewall.py | ActionGate | Combined gate: egress + firewall + guard | Middleware |
| circuit_breaker.py | TTLEnforcer | Session time-to-live enforcement | Logic |
| circuit_breaker.py | TokenCircuitBreaker | Token/cost/tool-call budget enforcement | Logic |
| circuit_breaker.py | StateRollback | Event-sourced action log with undo | Logic |
| hooks.py | SafetyGate | Tool blocking + rate limiting + VQC anomaly | Middleware |
| hooks.py | AuditLogger | SHA-256 attested action audit trail | Traceability |
| remediation.py | RemediationEngine | Auto-generated remediation playbooks from diagnostics | Prevention |
| config.py | HarnessConfig | YAML-driven harness configuration loader | Configuration |
| memory.py | MemoryManager | Active Forgetting memory with hash verification | Integrity |
The diagnostic finds what's broken. The playbook shows how to fix it. The SDK protects you continuously.
The NAIL Harness is the bridge between "we tested the agent" and "we continuously protect the agent." It's how you prove to underwriters that a software hallucination cannot translate into a catastrophic real-world loss.