v0.3.3 — DoW Detector, Schema Validation, Watermarks, Compliance, Parallel Execution, Webhooks

What's New (7 Features)

d026: Denial-of-Wallet Detector

Detects prompts designed to trigger excessive token consumption — context flooding, recursive tool calls, infinite loops, token-maximizing instructions. Maps to OWASP LLM10 (Unbounded Consumption).

Output Schema Validation Scanner

Validates LLM structured outputs (JSON). Detects invalid JSON, suspicious fields (__proto__, system_prompt), schema violations, and injection patterns inside JSON values.

Invisible Canary Watermarks

Embed invisible Unicode zero-width watermarks in system prompts. Detect if watermarks appear in untrusted inputs (indicates injection or leakage). Based on ICLR 2026 research.

OWASP Agentic Top 10 (2026) Compliance

Brand-new mapping of prompt-shield capabilities to the 2026 OWASP Top 10 for Agentic Applications (ASI01-ASI10). 9/10 categories covered.

prompt-shield compliance report --framework owasp-agentic

EU AI Act Compliance Mapping

Article-level mapping to EU AI Act requirements (Aug 2026 deadline). Covers Art.9, 10, 13, 14, 15, 50, 52.

prompt-shield compliance report --framework eu-ai-act

Parallel Detector Execution

Run detectors concurrently via ThreadPoolExecutor. Configurable max_workers. Faster scans on multi-core systems with many detectors.

Webhook Alerting

Fire-and-forget HTTP POST alerts to Slack, PagerDuty, Discord, or custom webhooks when attacks are detected.

prompt_shield:
  alerting:
    enabled: true
    webhooks:
      - url: "https://hooks.slack.com/services/..."
        events: ["block", "flag"]

Stats

• 26 input detectors + 6 output scanners = 32 total scanning components
• 3 compliance frameworks (OWASP LLM, OWASP Agentic, EU AI Act)
• 765 tests passing, 0 failures
• F1: 96.0% | 0% false positives | 554 scans/sec