Splyntra
Get demo
Security · Threat Defense & Risk Score

A real-time 0–100 risk score on every agent run

Security shouldn't be a post-mortem SIEM afterthought. Splyntra embeds inline prompt injection detection, PII/secret data loss prevention (DLP), and tool guardrails directly into the OpenTelemetry pipeline.

Start free on this feature →View documentationBook technical demo
risk_detector.py
python
# Runtime security inspection on agent spans
trace = splyntra.get_active_trace()

# Per-trace unified risk score (0 = clean, 100 = critical threat)
print(f"Risk Score: {trace.risk_score}") # -> 84/100 (HIGH RISK)
print(f"Threat Type: {trace.threat_type}") # -> "Indirect Prompt Injection"
print(f"Action Taken: {trace.guard_action}") # -> "BLOCKED_TOOL_CALL"
# Secrets & PII are redacted before data is stored
0–100
Unified Risk Score
Per-trace security rating mapped to exact spans.
< 5ms
Inline Detection
Real-time prompt injection evaluation.
100%
Zero PII Storage
Masks API keys, passwords, and sensitive PII.
3 Chokepoints
Runtime Defense
Model inputs, tool outputs, outgoing arguments.

Engineered for high-throughput autonomous agents

Every capability is built into the OpenTelemetry streaming pipeline with sub-millisecond ingestion overhead.

01Unified Risk

Unified 0–100 Risk Score Engine

Combines DLP violations, prompt injection probability, jailbreak patterns, and policy infractions into a single scannable score on every trace.

  • ✓Jump instantly from high-risk trace to the exact span that triggered it
  • ✓Configurable risk severity tiers (Low, Medium, High, Critical)
  • ✓Automatic alerts sent to security teams on elevated threat scores
02Threat Defense

Prompt Injection & Jailbreak Defense

Detects direct and indirect prompt injection inside user prompts, retrieved RAG documents, web scraping results, and MCP payloads.

  • ✓Inspects data at all 3 runtime chokepoints (input, retrieval, tool args)
  • ✓Identifies hidden instruction overrides, delimiter attacks, and ASCII smudging
  • ✓Inline tool-guard blocks unauthorized actions before execution
03DLP Engine

Automated Secret & PII Redaction

Scans all prompts, tool arguments, and model responses for API keys, SSH credentials, credit card numbers, and PII before storage.

  • ✓Deterministic pattern matching for OpenAI, AWS, GitHub, Stripe tokens
  • ✓Custom enterprise regex rules and entity redaction dictionaries
  • ✓Zero plaintext secret leakage to observability storage or dashboards
Architecture & Data Flow

How it works under the hood

Native OpenTelemetry ingestion, zero-overhead client sidecars, and real-time evaluation pipelines.

Step 1Ingestion

Payload Ingestion

Span text streams into the Splyntra OpenTelemetry collector.

Step 2DLP Guard

DLP & Secret Masking

Regex & ML scanners redact sensitive keys and PII in-memory.

Step 3Classifier

Prompt Injection Classifier

Lightweight embedding & heuristic models compute risk weight.

Step 4Policy Engine

Policy & Enforcement

If score exceeds threshold, block tool call or trigger alert.

Frequently Asked Questions

How is prompt injection detected without adding heavy latency?▼
Splyntra uses a dual-engine architecture: a high-speed sub-millisecond heuristic and structural scanner runs inline for immediate blocking, while asynchronous semantic classification scores the trace for forensics.
Can Splyntra block an agent from executing dangerous tools?▼
Yes. When paired with the Splyntra tool guardrail decorator or proxy, Splyntra evaluates risk before tool invocation and can abort high-risk execution paths.

Related Platform Features & Guides

GovernanceTrust Policies & Guardrails →GovernanceAudit Ledger →ComplianceCompliance Reports →

Ready to monitor and secure your AI agents?

Get started in under 3 minutes with zero credit card required. Free tier includes up to 5 projects and community telemetry.

Start Free on CloudRequest Custom Enterprise Demo