HUMΛN

Safety Protocols

Safety isn't aspirational—it's engineered. HUMΛN implements multi-layered safety systems to ensure AI agents operate within bounds, with human oversight, and fail-safe by default.

Safety Layers

Layer 1: Human-in-the-Loop

High-risk actions require explicit human approval before execution. The system identifies risk based on action type, delegated authority, and historical patterns.

python
# Automatically requires approval for high-risk tasks
workflow = client.humanos.orchestrate(
    task="Wire transfer: $50,000 to vendor",
    human_in_loop=True,  # Mandatory for financial actions
    required_capabilities=["financial_authority"]
)


# Workflow pauses until human approves
status = client.humanos.get_workflow(workflow.workflow_id)
print(status.status)  # 'pending_approval'

Layer 2: Capability Boundaries

Agents can only perform actions within their verified capabilities. Attempting to exceed capabilities results in immediate denial and provenance logging.

python
# Agent with 'invoice_processing' capability
try:
    client.humanos.orchestrate(
        task="Approve contract changes",
        required_capabilities=["contract_authority"]
    )
except HumanError as e:
    print(e.code)  # 'insufficient_capability'
    print(e.message)  # Agent lacks required capabilities

Layer 3: Delegation Constraints

Every delegation includes constraints: time limits, usage caps, spending limits, and allowed operations. The system enforces these automatically.

python
# Constrained delegation
delegation = client.delegation.create(
    delegator="passport_human",
    delegatee="passport_agent",
    scope=["expense_approval"],
    constraints={
        "expires_at": "2024-06-30T23:59:59Z",
        "max_uses": 100,
        "max_amount": 5000  # Per transaction
    }
)


# Agent attempts to exceed limit
try:
    approve_expense(amount=10000)  # Exceeds max_amount
except HumanError as e:
    print(e.code)  # 'delegation_constraint_violated'

Layer 4: Anomaly Detection

The system monitors for anomalous behavior: unusual usage patterns, rapid escalation attempts, or actions inconsistent with historical norms. Anomalies trigger automatic pauses and human alerts.

Example: Agent typically processes 10-20 invoices/day. Sudden spike to 200/day triggers automatic pause and notification to delegator.

Emergency Stop

Kill Switch Protocol

Humans can immediately halt any agent, workflow, or delegation at any time. This is not subject to override or delay.

python
# Immediately stop a workflow
client.humanos.emergency_stop(
    workflow_id="workflow_abc123",
    stopped_by="passport_human",
    reason="Unexpected behavior detected"
)


# Revoke all delegations for an agent
client.delegation.revoke_all(
    delegatee="passport_agent",
    revoked_by="passport_human",
    reason="Emergency suspension"
)


# All in-flight actions immediately fail
# Agent cannot take new actions
# Full audit trail preserved

Risk Classification

Risk LevelExamplesRequired Safeguards
LOWData retrieval, report generationCapability verification
MEDIUMDocument processing, data entry+ Delegation constraints
HIGHFinancial transactions, external APIs+ Human-in-the-loop approval
CRITICALLegal decisions, infrastructure changes+ Multi-party approval + Delay period

Audit & Compliance

Every safety event is logged and auditable:

Approval Decisions

Who approved/rejected what, when, and why

Constraint Violations

Attempted actions that exceeded delegation limits

Emergency Stops

Kill switch activations with full context

Anomaly Alerts

Detected anomalies and system responses

Related

Access Control

Capability-based permissions

Provenance Tracking

Audit trail for safety events