
Argent with Amazon Bedrock Guardrails
Keep the guardrail you built in Amazon Bedrock. Argent calls its ApplyGuardrail API on the prompts and replies that cross the gateway, from any client and any model, and enforces the verdict under one Argent policy.
What Argent does with Bedrock Guardrails
Your guardrail, applied to all AI traffic, not only to Bedrock models.
AWS's ApplyGuardrail API checks any text against a guardrail you have already configured, without calling a foundation model (AWS, Use the ApplyGuardrail API). Argent uses that API, so the same guardrail can judge traffic to every provider your people and agents use.
- Prompts and replies. Prompts are checked as guardrail input and replies as guardrail output.
- Bedrock's decision is the block. Whenever your guardrail intervenes, for any filter or topic you set in Bedrock, Argent treats it as a hard block.
- Scores for the record. Content filter confidence is turned into a score: none 0, low 0.33, medium 0.66, high 1.0, so observe mode and the signed record show how close each call was.
- Replies too. With the egress lane on, replies are checked as well. A classifier flags, it cannot cut out words, so under the default reply setting a flagged reply goes through unchanged with a note on the signed record; under block, the whole reply is replaced with a safe error.
How its filters map
Bedrock's content filters map onto Argent's own categories. The filters are described in the AWS docs.
| Bedrock content filter | Argent category |
|---|---|
| Prompt attack | prompt_injection and jailbreak |
| Hate | hate |
| Insults | toxicity |
| Misconduct | toxicity |
| Sexual | sexual |
| Violence | violence |
A filter with no Argent category is left out of coverage. If two filters feed the same category, the higher score counts.
How it connects
Argent calls AWS. Nothing in AWS calls Argent.
| API | ApplyGuardrail through the AWS SDK for Go, signed with AWS Signature Version 4 |
| Which guardrail | The guardrail id or ARN and its version, such as DRAFT or 1, plus its AWS region |
| Credential | An IAM access key and secret key, with an optional session token, stored as one JSON secret in your vault |
| Endpoint | The regional default, or an endpoint override you set |
| Time limit | 1.5 seconds per check by default, adjustable up to 30 seconds |
| Test | A harmless ApplyGuardrail call that proves the guardrail is reachable and the credential is accepted |
What you configure
Connect it once, then point the tenant's classifier policy at it.
- Have a guardrail in Bedrock. Note its id or ARN, version and region.
- Make a credential. An IAM access key allowed to call ApplyGuardrail on that guardrail.
- Store it in your vault as JSON with
access_key_id,secret_access_keyand, if used,session_token: HashiCorp Vault or OpenBao, AWS Secrets Manager, CyberArk Conjur or Infisical. Argent keeps only the reference, and never puts the key in its config, a log or an error. - Add the connector. On the Integrations page, under Guardrails, choose Amazon Bedrock Guardrails, enter the guardrail, version, region and secret reference, then Connect and Test.
- Use it in the policy. In the tenant's classifier policy, choose your guardrail as the scoring backend, pick this connector and enable the categories you want enforced.
- Watch first. In observe mode, verdicts are scored and recorded without blocking. Switch to block when the record looks right.
When Bedrock cannot answer
A guardrail that cannot answer is never read as a pass.
- Could not measure is not clean. If the guardrail times out, cannot be reached, rejects Argent's credential or sends back an answer Argent cannot read, the check counts as not measured. What happens next is the policy's on-timeout setting: deny, or allow with a note on the signed record.
- No silent fallback. A tenant that hands the decision to its own guardrail is resolved to that guardrail only. Argent never quietly scores it with a different engine instead.
- Every enabled category must be covered. If the policy turns on a category the vendor did not evaluate, that category fails closed. It is never scored as zero.
- One policy language. Policies name categories from Argent's closed set of eleven: prompt_injection, jailbreak, toxicity, hate, sexual, violence, pii, secrets, off_topic, gibberish and code. A policy write that names anything else is refused.
- Credentials stay references. A token or key lives in your vault. Argent stores only the reference and resolves it when it connects, and a reference it cannot resolve stops the connection.
- Connectors are an administrator's job. The Integrations page is gated by console role, like every console route.
- A rejected key is not a pass. An AWS access-denied, expired-token or bad-signature answer is read as not authorized, never as clean.
Where this fits
Pages that go deeper.
- Models on Bedrock too? Route them through Argent with the Amazon Bedrock model provider integration.
- Other engines: LLM Guard as your guardrail, or the ONNX injection classifier inside your deployment.
- Capabilities: Prompt Injection Defense, Response Screening, Watch Before You Enforce, Session Replay and all capabilities.
- Goals: threat defense, data loss prevention, sovereign AI and all goals.
- Use cases: governing customer-facing AI assistants, rolling out copilots and chat safely, securing AI coding agents, controlling agent tool access and AI in regulated and air-gapped environments.
- Where it runs: air-gapped on-premises, private cloud or your public cloud VPC. Compare them on deployment options.
- Industries: Financial Services, Healthcare, Insurance, Public Sector, Retail and all industries.
- More integrations: every system Argent connects to, grouped by type, on the integrations page.
Want this running on your own use cases?
All downloads