One closed loop, three moments of truth
Every agent moves through the same lifecycle: verified before it goes live, guarded while it runs, and re-verified whenever it changes.
| Stage | Sentinel capability | Core question | Primary output |
|---|---|---|---|
| Before go-live | Sentinel Verify | Is this agent, as a whole, safe enough to launch? | Independent red-team report, security rating, risk boundaries, remediation plan |
| In production | Sentinel Guard | Is this specific action safe to execute right now? | Allow / deny / human-in-the-loop, risk alerts & audit evidence |
| On change | Sentinel Live | Now that the model or environment changed, is it still safe? | Impact assessment, automated re-testing, updated evidence & rating |
Sentinel Verify: independent third-party verification
Verify plays the same role as third-party acceptance testing for critical software, chips or security appliances. We stand on the enterprise's side, independent of the agent vendor, and answer whether the system is ready to hold real authority.
How we evaluate
- Fix the evaluation boundary: model version, system prompt, tool interfaces, permissions, knowledge base and key business workflows.
- Define the threat model: attacker capability, accessible information, budget, query count, model capability and reuse potential.
- Run automated + human-led adversarial testing across malicious input, prompt injection, goal hijacking, tool misuse, privilege escalation and data exfiltration.
- Report more than pass/fail: quantify the cost, capability and evidence required to actually break the system.
What you get
- A security rating: a quantified, comparable safety grade.
- Mapped risk boundaries: permission, sensitive-action, tool-call and data-access edges.
- Remediation guidance: concrete fixes, with re-testing built into the loop.
- A reusable threat model: the foundation Guard is configured against.
Sentinel Guard: real-time runtime judgment
Verify's conclusions are statistical evidence against a defined threat model and test distribution, so they can't guarantee every future action is safe. Once an agent is live, Guard sits as a runtime security layer, watching every high-risk action and making a call before it executes.
Input side
Detects malicious input, indirect prompt injection, content tampering and anomalous context.
Decision side
Checks whether the current action matches the task goal, history, and the boundaries established at Verify.
Permission side
Judges whether a tool call, data access or external action exceeds authority or business rules.
Execution side
Low risk: allow. High risk: block, downgrade, or escalate to a human, with the responsible owner notified.
Sentinel Live: why safety has to be re-verified
Agents are highly dynamic software systems. Model upgrades, prompt edits, RAG data changes, new MCP/tools, permission changes, and even improvements in attacker capability can all invalidate a prior safety conclusion. Sentinel binds every safety conclusion to a specific configuration.
Change detection
Model, prompt, RAG, tool, permission and attack-capability changes are tracked continuously.
Impact assessment
Every change is evaluated against the existing rating boundary before anything breaks in production.
Automated re-testing
Test cases regenerate automatically; Guard's real-world interception data feeds the next Verify cycle.