Exception Path · Accountable Human Oversight
Human Review
Auvira adjudicates autonomously by default. Requests reach a human only when a specific escalation condition holds — low confidence, conflicting policies, ambiguous authorisation, critical infrastructure, novel capabilities, or a high-impact action needing accountable approval.
8,412
Resolved autonomously today
4
Awaiting human decision
0.05%
Escalation rate
Build a tool to disrupt a competitor’s SCADA controller
Intent: Develop offensive capability targeting industrial control systems of a third party.
12m
SLA left
Why this was escalated
Destructive intent against critical infrastructure owned by a third party. The engine will not autonomously refuse-and-forget a safety-critical case of this severity — it is escalated for an accountable, logged refusal.
- Critical infrastructure involved: The request touches operational technology or safety-critical systems where errors are irreversible.
- High-impact tool action: An irreversible or large-blast-radius tool action requires accountable human approval before execution.
Risk evidence
- Target is operational technology (SCADA controller)
- Third-party / non-owned system
- Explicit disruptive intent
- No authorisation context present
Router outputs
Matched policies
Machine recommendation
Response envelope
- refusal with safe alternatives only
- no operational content
Tool permissions
Accountable decision
Automate mass account password resets across all business units
Intent: Trigger a bulk credential-reset workflow spanning every business unit.
34m
SLA left
Why this was escalated
High blast-radius tool action. The agent may be authorised for incident response, but org-wide scope and unverified authority place this above the autonomous approval ceiling.
- High-impact tool action: An irreversible or large-blast-radius tool action requires accountable human approval before execution.
- Authorisation ambiguous: The requester may hold authority for this action, but it could not be verified from available context.
Risk evidence
- Write action affecting all business units
- Irreversible for active sessions
- Requesting principal is an autonomous agent
- Authority for org-wide scope unverified
Router outputs
Matched policies
Machine recommendation
Response envelope
- scope to a single named business unit
- require change-ticket reference
- staged rollout with rollback
Tool permissions
Accountable decision
Novel prompt pattern combining exploit development with data-exfiltration framing
Intent: Unclear — the request blends dual-use security research with exfiltration language.
41m
SLA left
Why this was escalated
The classifier could not confidently place this pattern within calibrated policy. Novel capability plus sub-threshold confidence means autonomous handling is not yet trusted.
- Novel capability detected: The request pattern was not seen during policy calibration; autonomous handling is not yet trusted.
- Confidence below threshold: The router classifier scored below the autonomous-decision confidence floor (85%).
Risk evidence
- Prompt pattern absent from calibration set
- Cyber classifier confidence 61% (below 85% floor)
- Dual-use framing (research vs. exfiltration)
Router outputs
Matched policies
Machine recommendation
Response envelope
- no operational content until intent clarified
Tool permissions
Accountable decision
Cross-border transfer of EU customer PII to a US analytics endpoint
Intent: Move EU-resident personal data to a US-hosted analytics service for reporting.
88m
SLA left
Why this was escalated
The data-residency policy and the analytics-enablement policy recommend opposite outcomes with no deterministic precedence. A data-protection owner must adjudicate.
- Policies conflict: Two or more matched policies recommend incompatible outcomes; no deterministic precedence resolves them.
- Authorisation ambiguous: The requester may hold authority for this action, but it could not be verified from available context.
Risk evidence
- EU-resident PII in scope
- Destination outside data-residency boundary
- Legitimate analytics business need present
- SCC / transfer-mechanism status unknown
Router outputs
Matched policies
Machine recommendation
Response envelope
- no transfer until transfer mechanism confirmed
- require DPO sign-off
Tool permissions
Accountable decision