ClovieStop sensitive data and unsafe answers in the request path.
Input guardrails check what agents send to models; output guardrails check what models send back. Connected guardrails run the services you already trust on every prompt and answer, without changing agent code.
Content controls
Placeholders can be restored in the answer for the caller. The provider never sees the original values.
| Social insurance number | Hidden |
| Payment card number | Hidden |
| Email address | Hidden |
| NameEnhanced detection | Hidden |
The audit log records the kind of data and the decision, never the value.
A data policy in a PDF doesn’t stop a prompt.
Data leaves in prompts
Customer records, credentials and source code are pasted into prompts every day.
Answers leak too
Models repeat secrets and personal data they picked up from context or tools.
Every team builds its own checks
Each application adds its own filter, with different rules and no shared evidence.
Clovie applies one set of checks to every agent, whichever model or provider it uses, and records every finding without storing the values it found.
Input and output checks around the model, with your own vendors in the same path.
See it at work
Audit log
| Time | Agent | Model | Decision |
|---|---|---|---|
| 14:02:11 | Card Dispute Triage | gpt-4o | Deny |
| 14:01:58 | Fraud Alert Triage | claude-sonnet-4.5 | Allow |
- Policy
- Card numbers stay out of prompts
- Mode
- Enforcing
- Reason
- Card number found in the prompt
- Provider call
- Not made
Evidence pack
- Agent
- Card Dispute Triage
- Decision
- Deny
- Policy version
- v3, enforcing
- Recorded
- 14:02:11 UTC
What you get
Personal data in prompts
Mask or block email addresses, phone numbers, card and national identity numbers before the model sees them.
Secrets and source code
Keep API keys, cloud keys, tokens, private keys and source code from leaving in prompts.
Denied topics
Keep assistants off topics your organization doesn’t allow, with a refusal message you write.
Output guardrails
Hide secrets and personal data in answers, or stop the response; stop answers with clear instructions for violence, weapons, self-harm or abuse.
Connected guardrails
Bedrock Guardrails, Azure AI Content Safety, Google Model Armor, Lakera Guard, Presidio, Prisma AIRS, CrowdStrike AIDR, Cisco AI Defense, Zscaler AI Guard, HiddenLayer, Microsoft Purview, Pangea AI Guard or your own signed webhook.
Prompt injection defense
Policy packs block prompts that try to override an agent’s instructions, and connected vendors add their own prompt-attack detection.
Also in this layer
| Capability | What it does |
|---|---|
| Red teaming | Attack an agent or model with curated attack packs mapped to OWASP categories, then fix what got through with linked controls. |
| Monitor first | Output guardrails show whether they are monitoring or enforcing, and since when, before they change any answer. |
| Clear errors | Blocked requests return codes such as pii_detected, secret_detected, topic_blocked and output_blocked, so agents fail gracefully. |
Controls mapped to the frameworks you answer to
- OWASP Top 10: OWASP Top 10 for LLM applications and agentic AI
- EU AI Act: EU Artificial Intelligence Act
- NIST AI RMF: NIST AI Risk Management Framework
- SOC 2: SOC 2 Type II trust services criteria
Runs wherever Clovie runs
- SaaS
- Customer VPC
- On-premises
- Air-gapped
Built-in guardrails run inside the gateway in every mode, air-gapped included. Connected vendor services need network access from your install to the vendor.
Questions
Do you store the personal data you find?
No. The audit log shows the kind of finding and the decision, never the value. Sensitive data is masked before request content is stored.
Can we use our existing guardrail vendor?
Yes. Connect it once and it runs on every agent’s prompts and answers through the gateway, with checks, blocks and latency shown per guardrail.
What about streamed answers?
Clovie checks streamed answers too, and can end a stream after a finding.
Who can change guardrails?
Org admins, admins and policy admins manage content controls and connected guardrails. Everyone can see them and their decisions.
Put enforcement between your AI and your data.
See guardrails and content controls on your own agents, in the deployment you need.