Skip to main content

How We Protect Keys in Your Prompts

If a governed agent prompt contains an API key, Clevername detects it, strips it before it reaches the AI model, and offers to store it securely in your BYOK key store. Available on every tier (Free, Pro, Team, Enterprise) with no configuration required.

Key Concepts

The short version

Every prompt routed through Clevername's governance gateway is scanned for API key patterns before it reaches any AI model. If a key is detected, it is immediately strippedfrom the prompt — the AI never sees it. You are then offered the choice to store it securely in your BYOK key store (GCP Secret Manager in production, AES-256-GCM in local dev) or discard it. This happens automatically on every tier with zero setup.

What we detect

Clevername recognizes API key formats from 30+ providers:

OpenAIsk-...
Anthropicsk-ant-...
Google AIAIza...
AWSAKIA...
AzureBearer ey...
Stripesk_live_...
GitHubghp_...
GenericLong random strings

Why this matters

API keys pasted into prompts are forwarded to AI providers in plaintext. This means your key could appear in provider logs, training data, or be exposed in a breach. Key interception prevents this by ensuring your key never leaves Clevername's backend.

What happens when you paste a key

  1. 1

    You send a message containing a key

    You paste "Use this key: sk-abc123..." into the chat. The message is sent to Clevername's backend over TLS.

  2. 2

    Key is detected and stripped

    Before the message is forwarded to any AI model, our scanner identifies the API key pattern and removes it. The AI receives your message with the key replaced by [KEY REDACTED].

  3. 3

    Provider is identified

    Clevername determines which provider the key belongs to (OpenAI, Anthropic, etc.) by matching its format. If the format is ambiguous, the key is validated against multiple providers in parallel to identify it.

  4. 4

    You are offered a choice

    The chat shows a message asking if you want to store the key securely. You can confirm (stores it in GCP Secret Manager), cancel (discards it), or ignore the prompt.

  5. 5

    Key is stored or discarded

    If you confirm, the key is encrypted and stored in GCP Secret Manager with CMEK protection — the same infrastructure used for keys you add manually. If you cancel, it is permanently discarded. Either way, the AI model never saw it.

Multiple keys per provider

You can store more than one key for the same provider. When a second key is detected for a provider you already have, Clevername asks you to give it a label(e.g., “Production” or “Testing”) so you can distinguish between them later.

Single key
OpenAI: sk-a...3f9b

Default setup. One key per provider, used for all requests.

Multi-key with labels
OpenAI (prod): sk-a...3f9b
OpenAI (dev): sk-x...7c2d

Label each key. Bind specific keys to specific agents for isolation.

Agent-key bindings

For teams running multiple AI agents, you can bind specific keys to specific agents. This means Agent A can only use the key you assigned to it — it cannot access keys bound to Agent B. Bindings are enforced at runtime by the guardrail engine.

AgentBound keyAccess
Support BotOpenAI (prod)Can use this key only
Code ReviewerAnthropic (dev)Can use this key only
Unbound agentNoneCannot use any key unless explicitly bound
Tip
Agent-key bindings are set during Agent Review approval. If your org uses agent review, the guardrail profile compiled from the questionnaire determines which keys each agent can access.

Key health monitoring

Clevername validates your stored keys every 6 hours by making a lightweight API call to the provider. If a key has been revoked, expired, or hit its rate limit, you are notified so you can replace it before your agents start failing.

Healthy

Key validated successfully. Provider returned a valid response.

Warning

Key is valid but approaching a rate limit or usage cap.

Invalid

Key was revoked, expired, or rejected by the provider. Replace it.

How key interception is audited
1

Detection logged

Every key detection event is logged with the provider, whether it was stripped, and the action taken (stored, discarded, or held). The key value itself is never included in logs.

2

Storage logged

If you confirm storage, the write to GCP Secret Manager generates a Cloud Audit Log entry with a timestamp and the service account that performed the write.

3

Usage logged

Every time a stored key is fetched for an API call, a DATA_READ audit log entry is created. Enterprise customers can export these to their own SIEM for independent verification.

SOC dashboard visibility

The SOC console includes a Key Interception card showing 24-hour stats: keys intercepted, keys stored, keys rotated, and keys denied. Admins can see trends and respond to anomalies (e.g., a sudden spike in detected keys may indicate a credential leak).