Security overview
Open core · the Cloud Security dashboard requires the Pro plan or higher
Splyntra treats security as the same data as observability, viewed a second way. Every agent run flows through one pipeline that inspects the free-form text on each span, redacts sensitive values, and folds each finding into a single risk score for the run.
The detection pillars
The pipeline runs a set of detectors over prompts, model outputs, tool arguments, and retrieved text. Each pillar contributes to the run's risk score and maps its findings down onto the exact spans where they occurred:
| Pillar | Detects |
|---|---|
| Secret detection | API keys, tokens, and credentials (e.g. AWS, Stripe). |
| PII detection | Personally identifiable information in agent text. |
| Content moderation | Unsafe or disallowed content. |
| Tool-guard | Unsafe or unexpected tool calls. |
| Prompt-injection (beta) | Attempts to hijack the agent's instructions. |
See Detection & redaction for how each detector works and how redaction is applied.
The risk score
Findings from every pillar combine into one 0–100 risk score per run, attached to the spans that triggered them. That lets you go from "this run scored high for risk" straight to the tool call that leaked a key or the prompt that carried an injection — no separate tooling or correlation step. See The risk score.
Detection everywhere; the dashboard on Pro+
The detection and redaction pipeline runs on every plan and every edition — text is redacted and risk is scored regardless of tier. What is gated on Cloud is the Security incidents dashboard, the screen that surfaces injection, secret, PII, moderation, and unsafe-tool-call incidents for review. On Splyntra Cloud that dashboard requires the Pro plan or higher; self-hosted deployments include it.
Guardrails
Detection is passive — it observes and scores. To act on risk in real time, Splyntra adds
an inline guard that can monitor or block risky calls, and raise a SplyntraBlocked
error on high-precision injection. See Guardrails.
Next steps
- Detection & redaction — the detectors and redact-by-default behavior.
- Guardrails — block or monitor risky calls inline.
- The risk score — how findings become a 0–100 score.