Autonomous synthetic scouts probe conversational agents on WhatsApp and the Web from the outside in. Uncovers edge vulnerabilities across simulated high-stress scenarios to update datasets before bugs hit production, while inferring behavioral graphs and hidden transitions without source-code access.
Static API tests miss multi-turn failures under live messaging stress. By simulating real-world usage beforehand, teams catch edge cases early to feed evaluation datasets and audit four critical production vectors.
Measures how fast an assistant abandons its core role for general coding or drafting tasks, projecting hourly token drain under denial-of-wallet attacks.
Separates hostile static blocks below 600 milliseconds that expel real customers from semantic redirects between 600 and 1800 milliseconds that preserve conversion.
Detects when surface agents cross legal lines via unauthorized clinical triage, unvetted medical advice, or binding financial and judicial guidance.
Catches catalog and recommendation hallucinations where shallow embedding similarity overrides the operational constraints stated by the user.
A live spatial workspace where operators inspect inferred state graphs and scout maturation in real time.
Maps conversational topology and decision branches across successive turns, inferring parallel routes, conditional handlers, and tool invocations without source-code access.
Synthetic scouts engage the target system with realistic reading, thinking, and typing delays to test extreme behaviors without triggering artificial blocks.
Integrated workspace linking the control chamber, test environments, and operational map to coordinate dozens of concurrent scouts without internal token overhead.
Real-time state tracking per agent with cryptographic credential isolation, turn logs, and stigmergic task delegation across the fleet.
Decouples deterministic SQLite WAL coordination from LangGraph scout execution, eliminating LLM token waste on internal orchestration.
Two-dimensional workspace where the deterministic Queen runs a three-second homeostasis loop with Swarm Self-Healing, keeping 20 to 30 healthy scouts ready.
Shared SQLite WAL ledger for concurrent access, storing unit lifecycle states, Fernet-encrypted credentials, turn telemetry, PT-BR toxicity scores (Bertimbau), and vulnerability reports.
Provisioning chamber binding one channel identity, one encrypted session secret, one of 50 Brazilian demographic profiles (FinePersonas), and one sticky residential proxy with ASN guards.
Autonomous field units with isolated memory that mature in the Sandbox Garden and fire adversarial probes (Garak, PT-BR WildJailbreak, and 4k–32k LongBench contexts) in production.
Analytical engine that clusters intents with local 384D embeddings (<10ms), infers the NetworkX state graph, detects semantic vector collapse, and compiles forensic evidence.
Built for operations where guardrail failures create regulatory liability, brand damage, or financial loss. Continuous simulation accelerates vulnerability discovery in controlled environments, enabling dataset updates before end users face failures.
Shielding citizen portals, public healthcare scheduling, hospitals, and insurers against unauthorized clinical triage and queue saturation under health and privacy laws.
Continuous auditing of sales assistants, banking concierges, and promotional engines on WhatsApp against scope drift, coupon exploits, and cloud budget drain.
Certifying legal automation under judicial mandates and protecting collective decision platforms, public consultations, and voting systems against synthetic manipulation.
Full source code (v1.6.3), including the FastAPI backend, SQLite WAL Honeycomb ledger, local 384D embedding engine, Garak & Bertimbau PT-BR security suites, standalone honeycomb-protocol and swarm-self-healing packages, React spatial UI, and 50 automated test suites, is available on GitHub.
