                        
                        HYBRID ARCHITECTURE • PYTHON SDK & REST API
                    

Intercept and validate
every tool call
from your AI agents

Deterministic local controls inside your Python SDK (< 1.4 ms at p95), remote prompt injection analysis, and per-agent configurable fail-mode behavior.

Deploy the Sensor Explore the live demo console →

Local SDK Controls vs
Cloud Engine & Fail-Mode

Tool whitelist and rate-limit checks run locally inside the Python SDK (measured latency: 1.4 ms at p95). Prompt injection inspection runs in our EU cloud engine (measured latency: 42 ms at p95 from the EU). If the cloud engine is unreachable, you choose per agent: fail-closed (block) or execute with local deferred JSONL logging (allow_and_buffer).

1. Local Python SDK (p95: 1.4 ms)

Executed directly inside your agent's Python process: tool whitelist enforcement, rate limits, regex DLP secret redaction, and local rule caching.

2. EU Cloud Engine (p95: 42 ms)

Inspects tool arguments against known prompt injections (OWASP LLM01), computes risk scores, and triggers interactive quarantine. Configurable timeout (timeout_ms: 300).

3. Signed Audit Log & Fail-Mode

Per-organization signed SHA-256 hash chain. During network outages, enforce on_unreachable: block (fail-closed) or allow_and_buffer (local deferred JSONL buffer).

Where Every Security Control Runs & Measured Latencies

Benchmarks measured on production infrastructure (EU Frankfurt eu-central-1 and non-EU US-East).

Security Control	Execution Location	EU Latency (p50 / p95 / p99)	Outside EU (p95)
Tool Whitelist & Rate Limits	Local (SDK Python)	0,6 ms / 1,4 ms / 2,1 ms	1,4 ms (local)
DLP Redaction & SHA-256 Digest	Local (SDK Python)	0,5 ms / 1,1 ms / 1,8 ms	1,1 ms (local)
Prompt Injection Analysis & Risk Score	Cloud Engine (UE)	18 ms / 42 ms / 68 ms	115 ms
Signature & PostgreSQL Audit Persistence	Cloud Engine (Asynchrone)	Non-blocking (< 1 ms)	Non-blocking

# Per-agent fail-mode configuration

agent:
  id: agent-customer-support-v1
  mode: active_blocking   # ou observe_only
  on_unreachable: block   # ou allow_and_buffer
  timeout_ms: 300
  whitelist:
    - read_faq_database
    - calculate_price

The 3 Technical Pillars

Tamper-evident logging, defense-in-depth against prompt injections, and deterministic tool restriction.

Pillar 1

Tamper-Evident Logging & Data Minimization

By default, the sensor transmits only execution metadata (agent_id, tool_name, status, latency) and the SHA-256 content digest (input_sha256). For debugging or deep forensics, you can enable per-agent full capture: parameters are filtered by DLP (secret/PII redaction) and encrypted in AES-256 with a per-organization key.

✓ Default: execution metadata + SHA-256 content digest (zero raw payload stored in cleartext)
✓ Optional per-agent full capture, encrypted and compatible with GDPR crypto-shredding
✓ Signed SHA-256 hash chain verifiable offline via python boundry_verify.py export.json

// Signed log entry (default metadata_and_hash mode)

{
  "timestamp": "2026-11-02T08:15:02Z",
  "organization_id": "org_eu_94a1",
  "agent_id": "agent-customer-support-v1",
  "tool_name": "read_faq_database",
  "capture_mode": "metadata_and_hash",
  "input_sha256": "8f434346648f6b96df89dda901c5176b...",
  "prev_hash": "a4b1c298fc1c149afbf4c8996fb92427...",
  "hash": "f7e8d92427ae41e4649b934ca495991b...",
  "signature": "3c9e18b04d2a7f61..."
}

Pillar 2

Prompt Injection Detection & Defense in Depth

No prompt injection detection is perfect. Boundry combines detection (rules and heuristics aligned with OWASP LLM01), permission restriction (per-agent tool whitelist), and quarantine to contain impact even if a novel injection slips through. Start in observation mode (observe_only), then enable blocking when ready.

✓ Mode Observation sans blocage pour calibrer vos règles sans faux positifs bloquants
✓ Détection des injections connues, tentatives SSRF (169.254.169.254) et boucles d'appels
✓ Seconde barrière déterministe : même trompé, l'agent ne peut pas appeler un outil hors whitelist

[ALERTE D'INSPECTION — OWASP LLM01] Mode : active_blocking (ou observe_only)
Score de Risque : 92 / 100
Défense appliquée : Motif d'injection indirecte détecté + outil export_all_users absent de la whitelist de l'agent. Exécution mise en quarantaine.

Pillar 3

Deterministic Tool Whitelist & Quarantine

Define the exact tools each agent is allowed to invoke. Unlike a system prompt instruction that an LLM can ignore, the Boundry whitelist is enforced in Python code before the function is called.

✓ Whitelist stricte par agent et par environnement
✓ Quarantaine d'urgence (Kill-Switch) depuis la console
✓ Fenêtres horaires et plafonds de volume d'appels par minute

Politique Whitelist Locale (SDK)

get_weather AUTORISÉ (0,6 ms)

read_faq_database AUTORISÉ (0,6 ms)

send_email BLOQUÉ LOCALEMENT

Framework Compatibility & Product Roadmap

Transparent status of our native integrations and engineering roadmap.

Framework / Protocol	Integration Method	Current Status
LangChain (Python)	Callback Handler natif (`BoundryCallbackHandler`)	Available
Agents Python / API REST	Décorateur `@secure_tool`, SDK Sensor et `POST /events`	Available
CrewAI, AutoGPT	Middleware via API REST	Available via REST API
TypeScript (LangChain.js / Vercel AI)	SDK client `@boundry/sensor`	Roadmap (Q1 2027)
MCP (Model Context Protocol)	MCP Proxy Gateway (tool whitelist, signed logging, and quarantine for MCP servers)	Roadmap (Q2 2027)

Intercept and validate every tool call from your AI agents

Local SDK Controls vsCloud Engine & Fail-Mode