VaultCall

🛡️ VaultCall: Autonomous Voice AI Protocol for BEC Wire Defense 💡 Inspiration: The $55 Billion Crisis According to the FBI Internet Crime Complaint Center (IC3), Business Email Compromise (BEC) is the single most damaging cybercrime category in history, with cumulative enterprise losses exceeding $55 Billion.

The attack vector is deceptive and devastating: an attacker compromises a vendor's corporate email inbox, monitors open billing conversations, and sends an urgent message with altered bank routing numbers right before payment settlement. Accounts Payable (AP) specialists and treasury clerks—handling dozens of invoices daily under strict deadlines—routinely release six- and seven-figure wire transfers directly to criminal mule accounts.

SOX 404 internal control guidelines explicitly mandate: "Never verify bank account changes over email; always conduct an out-of-band telephone call to a known corporate contact."

However, human phone verification fails in practice:

AP staff are overwhelmed: Waiting on hold for 40 minutes per vendor creates unacceptable payment bottlenecks. Social engineering & burner numbers: Attackers embed their own burner phone numbers in fraudulent PDF invoice footers, tricking AP clerks who dial the number on the paper. Corner-cutting: Under quarter-end pressure, clerks skip the call entirely. We asked: What if an autonomous Voice AI agent could execute an airgapped, cryptographic telephone challenge-response protocol directly with corporate officers over PSTN, verify their tax credentials, and issue an irrevocable audit certificate before a single dollar leaves the bank?

That vision became VaultCall.

⚙️ What It Does VaultCall sits directly between enterprise ERP systems (SAP, Oracle Treasury, Workday) and outbound wire settlement rails (Fedwire, SWIFT). Whenever a vendor bank modification or high-value wire is triggered, VaultCall initiates an automated defense sequence:

Airgap Policy Gate: Intercepts and incinerates any phone number found on incoming emails or invoices. Outbound calls are strictly routed to hardened corporate PBX directory numbers. Dynamic Challenge-Response Token: Generates single-use NATO phonetic challenge tokens (e.g., $\text{Token} = \text{Echo-Sierra-482}$) and demands verification of the last 4 digits of the vendor’s Federal Tax ID (EIN). Live Telephony Dialogue via CALL-E: CALL-E dials the officer over standard telephone networks. It conducts an intelligent, natural dialogue—handling interruptions, background noise, identity inquiries, and multiple languages (including English and Hindi). Evidence-Linked Scorer: Verbatim utterances are cross-examined against extracted claims using an evidence confidence function: $$\text{Score}(V) = \sum_{i=1}^{n} w_i \cdot \mathbb{I}(\text{Claim}i \in \text{Utterance}{\text{verified}}) - \lambda \cdot \mathbb{I}(\text{FraudSignal})$$ Any unevidenced or evasive response fails closed, automatically freezing the transaction. Irrevocable SHA-256 Voice Certificate: Approved transactions generate a cryptographic audit receipt with a SHA-256 fingerprint: $$\mathcal{H} = \text{SHA-256}(\text{VendorID} \parallel \text{Timestamp} \parallel \text{Token} \parallel \text{AudioDigest})$$ This furnishes auditors with indisputable voice proof for SOX 404 and SOC2 Type II compliance. One-Click Enterprise Kill Switch: Allows treasury directors to instantaneously freeze all outbound dialing and pending settlements globally in the event of an active cyber incident. 🛠️ How We Built It VaultCall was architected from the ground up as a zero-trust financial defense application:

Voice AI & Telephony Engine: Integrated CALL-E's REST API and Webhook infrastructure to initiate real-time PSTN voice calls, stream low-latency audio, and capture conversational transcripts. Full-Stack Dashboard: Built with Next.js 14 App Router, TypeScript (strict mode), and TailwindCSS. The interface includes an executive command center, real-time waveform visualizers, side-by-side invoice diff viewers, and forensic audio audit reels. Airgap & Security Layer: Designed a zero-trust registry lookup (airgap-gate.ts) that cross-references incoming requests against verified PBX phone records, preventing spoofed caller-ID and burner-phone redirects. Cryptographic Audit Pipeline: Implemented SHA-256 digital certificate generation (evidence-scorer.ts and store.ts) that binds call metadata, verbatim proofs, and ERP release tokens into tamper-evident audit receipts. Deterministic Offline Fixtures: Created a deterministic test harness (seed-data.ts and vitest) with 14 automated unit and integration tests, allowing hackathon judges to verify edge cases (clean confirmation, active fraud interception, and gatekeeper/receptionist handling) instantly with zero API friction. 🧗 Challenges We Faced Combating Invoice Number Spoofing: In real-world BEC attacks, hackers replace the vendor's phone number on the invoice with their own burner cell phone. If an AI calls whatever number is in the payload, the hacker answers and approves the wire! We resolved this by building the Airgap Policy Gate, enforcing an architectural separation: invoice amounts are read from the transaction, but phone numbers are never accepted from untrusted sources. Handling Natural Human Speech in Live Calls: Real executives don't speak like robots. In live calls, officers interrupt, ask "Who is this and how did you get my number?", ask the agent to wait, or switch languages. We engineered CALL-E's prompt compilation to preserve conversational context, maintain calm professionalism, answer security credential questions, and verify multi-lingual responses without hallucinating. Evidence-Linked Scoring & Anti-Hallucination: Financial security cannot rely on fuzzy LLM summaries. If an agent hallucinated an approval, an enterprise could lose millions. We enforced a strict mathematical evidence-linking model where every verdict requires verbatim citation back to the callee's transcript. 🏆 Accomplishments That We're Proud Of Real PSTN Telephony Tested in the Real World: Successfully placed real phone calls over actual telecom networks to international numbers, conversing fluidly in Hindi and English, proving that CALL-E functions reliably in live production conditions. Intercepting a $785,000 Cyberattack: Demonstrated a complete end-to-end simulation of an active BEC attack where an attacker’s burner phone was intercepted, the real controller was reached, fraud was detected, and the wire was locked down in real time. 100% Test Suite & Judge Fast-Path: Designed 14 comprehensive Vitest test suites that pass completely offline, ensuring zero friction for anyone cloning and evaluating the repository. Enterprise-Grade Visuals: Created a polished, dark-mode executive console with real-time waveform animations, forensic diffs, and SOX 404 audit certificates. 📚 What We Learned Voice is the Missing Layer in Zero-Trust Security: While email, SMS, and push notifications are increasingly vulnerable to phishing and session hijacking, real-time voice verification over PSTN combined with dynamic challenge tokens provides a resilient, out-of-band defense. The Power of Low-Latency Telephony: CALL-E's ability to handle bidirectional voice conversations with human-level naturalness makes autonomous enterprise workflows possible for high-stakes decisions that previously required human AP staff. 🚀 What's Next for VaultCall Direct ERP Webhook Connectors: Pre-built integration plugins for SAP S/4HANA, Oracle NetSuite, and Workday Financial Management. Biometric Voiceprint Verification: Integrating acoustic voiceprint matching with CALL-E to verify speaker cadence and frequency against previous enrollment calls. Automated Fedwire / SWIFT Settlement Hooks: Directly interfacing with ISO 20022 banking APIs to automatically release approved wires upon cryptographic certificate generation.