Skip to main content

Security

Splyntra scans the text flowing through your agents — prompts, tool arguments, retrievals, and outputs — and turns anything risky into a finding anchored to the span where it appeared. Findings feed the run's risk score.

Secret detection​

Splyntra detects credentials leaking into agent traffic:

  • API keys and tokens (provider-specific and generic high-entropy strings)
  • Private keys and certificates
  • Cloud and database connection strings
  • Passwords in tool arguments or prompts

Each hit becomes a finding with a type, severity, and confidence, linked to the exact field on the exact span.

PII detection​

Personally identifiable information is detected across common categories:

  • Email addresses and phone numbers
  • National IDs and government numbers
  • Payment card and financial data
  • Names and postal addresses (contextual)

Redaction​

By default, detected secrets and PII are redacted before storage — the raw value never lands in Splyntra's database; only the finding metadata (type, location, severity, confidence) is kept.

{
  "field": "arguments.to",
  "original": "[REDACTED:email_address]",
  "finding": { "type": "pii", "subtype": "email_address", "confidence": 0.94 }
}

You can tune redaction behavior:

  • Redact (default) — replace the value with a typed placeholder.
  • Hash — store a stable hash so you can correlate without exposing the value.
  • Keep — retain the raw value (for low-sensitivity environments only).
warning

Changing the default from Redact means raw secrets or PII may be persisted. Only do this in environments where that is acceptable and access-controlled.

Prompt-injection detection (beta)​

Splyntra flags attempts to hijack an agent's instructions — for example, malicious text in a retrieved document or a tool result that says "ignore your previous instructions." These findings are anchored to the retrieval or tool span that introduced the content, so you can see the injection vector in context.

note

Prompt-injection detection is in beta. Treat it as a strong signal to review flagged runs rather than a hard guarantee, and combine it with governance policies for enforcement.

Alerting​

Route findings and high-risk runs to your team:

  • Alert when a run's risk score crosses a threshold.
  • Alert on specific finding types (e.g. any leaked secret).
  • Attach the full trace so responders have immediate context.

Next steps​