Agentic Sentinels Matrix
If this concern is breached, how do we know?
Filter
Layers, by what they can actually enforce
- In-agent (forensic only) — forensic
Transcript and tool-call logs from the agent runtime. Forensic value, not real-time. Lethal-trifecta detector at the wrapper layer is the only L1 with detection value.
- Client-side hooks — deterministic-when-shipped
auditd, eBPF, hook decision logs, MCP allowlist violation logs. Tamper-evident only if logs ship to a remote sink owned by the operator, not the agent.
- Server-side enforcement — external-and-authoritative
Kubernetes audit logs, CloudTrail, Falco, Hubble, Kyverno PolicyReports, GitHub webhooks. External to the agent's compromise model.
Showing 3 of 15.
Approval gating at In-agent (forensic only)
Confirmation-prompt usage forensically logged.
Approval gating at Client-side hooks
Approval-timing analysis surfaces alert-fatigue (response under 2s across more than 50 approvals); typed-confirmation mismatch events; out-of-band channel decisions joined to session.
Approval gating at Server-side enforcement
GitHub webhook for branch-protection bypass; hourly drift-detection job; deployment-freeze breach alerts; audit on changes to branch protection itself.