Data security for AI
The fastest way to leak data through AI is to paste it into a prompt. 18 of 56 vendors document controls that detect, redact or map sensitive data in prompts, responses, training data and agent workflows. Fortinet FortiAIGate states 35 PII and confidential-data types in its detector. Cloud providers build it into the model path: Amazon Bedrock Guardrails, Google Model Armor and Microsoft.
18 vendors that document data protection for AI
Checked 11 Nov 2026| Vendor | What it documents | Also covers |
|---|---|---|
| Akamai Firewall for AI | Inspects model responses for PII leaks and intellectual-property disclosure. 1 | Guardrails |
| Alice (ActiveFence) | Redacts PII, credentials, internal system details, and other sensitive information. 1 | Guardrails, Red teaming |
| Amazon Bedrock Guardrails | Blocks or masks PII and custom regex matches in inputs and responses. 1 | Guardrails |
| Aurascape | Uses real-time sensitive-data classification and fingerprinting. 1 | Guardrails, Agent security, Red teaming, AI-SPM |
| Cisco AI Defense | Controls employee access to third-party AI applications. 1 | Guardrails, Red teaming, AI-SPM, Model scanning |
| Cloudflare AI Security for Apps | Detects PII in incoming prompts including contact, financial, and identity information. 1 | Guardrails |
| Coralogix AI Guardrails | Security guardrails prevent data and PII leaks and sanitize inputs. 1 | Guardrails, Agent security |
| Cyera AI Guardian | Checks prompts against identity and intent to keep data access authorized and stops sensitive data reaching public models. 1 | Agent security, Red teaming, AI-SPM |
| Fiddler AI | Detects and redacts PII, PHI, and secrets before model input and response delivery. 1 | Guardrails, AI-SPM |
| Fortinet FortiAIGate | DLP detects 35 PII/confidential-data types, redacts responses, and blocks tool-call leaks. 1 | Guardrails, Agent security |
| Giskard | RBAC, audit trails, identity-provider integration, and a zero-training policy. 1 | Agent security, Red teaming |
| Google Cloud Model Armor | Scans and redacts sensitive IP and PII from prompts and responses. 1 | Guardrails |
| IBM Guardium Exposure Manager | Alerts on risky sharing, AI-generated outputs, and policy violations involving sensitive data. 1 | AI-SPM |
| Microsoft AI security | Detects known protected text material in AI-generated responses. 1 | Guardrails, Agent security, AI-SPM |
| Orca Security | Scans AI projects and model-training data for sensitive information. 1 | AI-SPM |
| Proofpoint AI Security | Applies controls to prompts, responses, tool calls, and workflows. 1 | Guardrails, Agent security, AI-SPM |
| Upwind | Detects sensitive data in prompts and inference payloads in real time. 1 | Agent security, Red teaming, AI-SPM |
| Wiz | DSPM capabilities identify sensitive training data and exposure paths. 1 | Guardrails, AI-SPM |
What to ask before you shortlist
- Where does it act? At the prompt, at the response, on the training set or at rest in data stores. Each needs a different product. IBM Guardium Exposure Manager maps data connected to RAG pipelines and vector databases.
- Redact or block? Redaction keeps the workflow running; blocking stops it. Check which the vendor documents.
- How many data types? Few vendors publish a detector count. Ask for the list and for custom pattern support.
- Where is your data processed? Giskard documents a choice of EU or US processing. Compare deployment on each vendor page, or filter the table for self-hosted options.