Quarantine
Quarantine is a safety mechanism in Synentra that temporarily isolates agents exhibiting suspicious, harmful, or policy‑violating behaviour. Once quarantined, an agent cannot make further requests until a human operator reviews and releases it (or an automatic expiry kicks in).
Quarantine acts as a circuit breaker for agent behaviour – it stops a potentially compromised or misbehaving agent from causing damage, while giving operators time to investigate.
Why Quarantine is Essential
Autonomous agents can go rogue for many reasons:
| Cause | Example |
|---|---|
| Compromised credentials | An attacker steals an agent's JWT and uses it to exfiltrate data |
| Bugs in agent logic | An agent enters a loop, making thousands of requests per second |
| Policy violations | An agent repeatedly attempts unauthorised actions |
| Anomalous behaviour | An agent starts acting outside its normal pattern (e.g., off-hours traffic) |
Without quarantine, these issues would go unnoticed until significant damage occurs. Quarantine provides an automatic safety net.
How Quarantine Works
Quarantine Lifecycle
1. Detection
An agent enters quarantine when one of the following conditions is met:
- Trust Score Falls Below Threshold – Each agent has a dynamic trust score (0.0 – 1.0). When it drops below the configured floor (default: 0.3), the agent is automatically quarantined.
- Excessive Violations – If an agent repeatedly triggers
Denydecisions within a short time window, it may be quarantined. - Manual Intervention – An administrator can manually quarantine an agent via the admin API or CLI.
2. Isolation
When quarantined:
- All requests are blocked – The agent receives a
403 Forbiddenresponse for every subsequent request. - No further evaluation – The policy engine and risk scoring are skipped entirely, saving resources.
- Active sessions are invalidated – Any JWT issued for the agent may be revoked (optional).
3. Review & Action
Human operators can:
- Release – Restore the agent to
Activestatus, allowing it to resume operations. The trust score remains low but will increase over time if behaviour improves. - Revoke – Permanently disable the agent. Revocation is irreversible and may trigger additional alerts.
- Leave Quarantined – If the issue is being investigated, the agent can remain quarantined until a decision is made.
5. Automatic Release (Optional)
If configured, an agent may be automatically released from quarantine after a cooldown period (e.g., 1 hour) or when its trust score recovers (e.g., after successful low‑risk requests). This is useful for transient issues like temporary spikes in request rate.
Quarantine vs. Revocation
| State | Description | Reversible? |
|---|---|---|
| Active | Normal operation – agent can make requests | – |
| Quarantined | Temporarily blocked, pending review | Yes (can be released) |
| Revoked | Permanently deactivated – agent cannot be used again | No |
💡 Tip: Use quarantine for suspected issues that can be resolved (e.g., misconfigured agent). Use revocation for agents that are definitively compromised or no longer needed.
Benefits of Quarantine
| Benefit | Description |
|---|---|
| Rapid Incident Response | Automatically stops an active threat before it can cause damage |
| Reduced Mean Time to Respond (MTTR) | Operators are immediately alerted and can act |
| Protection Against Compromised Agents | Limits the blast radius of a compromised credential |
| Policy Enforcement | Agents that repeatedly violate policies are quickly isolated |
| Operational Safety | Gives teams time to investigate without pressure |
| Automatic & Manual Control | Balances automation with human judgment |
Quarantine is a critical safety layer in Synentra's governance model. It ensures that even when agents go wrong, they are contained quickly and with minimal disruption – giving you the confidence to deploy autonomous agents at scale.