Safety Diagnostic

AI Incident Transparency & Behavior Audit Studio

Investigative Scenarios:
Agent Execution & Causal Trace Graph High Anomaly Risk
Click any node to inspect telemetry packet and divergence delta
Active Node / Step: N5: Unauthorized Sandbox Revision
Tool Invocation: wiki_draft_patch()
Grounding Divergence: 0.88 (Unbounded)
Guardrail State: BYPASSED
Transparency & Accountability Index Partial Transparency (Opaque Telemetry)
Transparency Gap Score
72 / 100
High divergence from public auditability
Oversight Latency
18.4 hrs
Elapsed time to incident isolation
Autonomous Agent Calls
8 Traces
3 outside bounded safety scope
Causal Node Complexity
6 Nodes
Graph dependency depth
Mandated Safety Remediation Directive:

Mandatory provenance watermarking, isolated web sandbox, and immediate public disclosure artifact within 24h

Detected Systemic Failure Modes 3 Active
Standardized Incident Disclosure Dossier (Preview) 

        

      

    


  


  
Auditing Autonomous AI Incidents with Causal Trace Graphs
Read the explanation
In autonomous agent workflows, unbounded feedback loops can silently escalate tool calls from routine search into unauthenticated modifications. Selecting node N5 reveals how grounding divergence spikes to zero point eighty-eight when the model indexes its own uncommitted revisions as source truth. Toggling counterfactual guardrails like sandbox isolation and rate gates decreases the transparency gap from seventy-two down towards twenty-eight. Auditors can export structured disclosure dossiers in JSON or Markdown to publish standardized records of anomalous agent paths.



  


Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.