- A CFO joined a video call with his own team. He approved $25M in transfers. Except it wasn't him. Every person on that call was an AI-generated deepfake. The Arup breach rewrites everything we thought we knew about trust. 🧵
- February 2024. Arup's Hong Kong office. An employee joins a video conference with the CFO and several senior executives. Familiar faces. Familiar voices. He follows their instructions and wires $25.6M across 15 transactions to five bank accounts.
- Every single person on that call was synthetic. Generated in real-time by GANs trained on publicly available YouTube talks and conference footage. Arup's network was never breached. No malware. No credential theft. They hacked something deeper: human trust.
- The attackers didn't bypass security. They bypassed reality. They used video injection — feeding AI-generated streams directly into the conferencing software. The app treated synthetic packets as a real camera feed. No liveness check could catch it.
- This is the new attack surface. Face swap attacks rose 704% in 2023. Injection attacks on ID verification jumped 255%. The cost to generate a hyper-realistic deepfake twin? About $15 and 45 minutes of effort. The perimeter defense model is dead.
- Here's what most orgs miss: their "AI strategy" actually makes this worse. LLM wrappers send sensitive data — financial records, exec comms — to third-party clouds for processing. You're defending a castle while shipping your crown jewels out the back door.
- Wrappers are probabilistic. They predict tokens, not truth. When your AI agent confirms a price or authorization, it's guessing based on statistical patterns — not querying your actual database. In high-stakes moments, that gap is where fraud lives.
- Our answer: Deep AI. Private LLMs deployed inside your VPC. A neuro-symbolic "sandwich" — deterministic logic on both sides of the neural network. The model reasons. The logic layer verifies against ground truth. No guessing. No data egress.
- For identity, we need to go beyond what eyes and ears can verify. Physiological signals like heartbeat-induced micro-color changes in skin. Behavioral biometrics — keystroke cadence, mouse dynamics. These are nearly impossible to forge in real time.
- The real shift is cultural. "Verify first" must replace "comply immediately." Out-of-band confirmation for any high-value instruction. Dual authorization from someone NOT on the original call. Reward skepticism — even toward the C-suite.
- When a CFO's face and voice can be fabricated in minutes, what should actually count as "proof of identity" inside an enterprise? #DeepfakeDefense #EnterpriseAI
We broke down the full forensics, the technical architecture, and the resilience roadmap in our latest whitepaper: - https://veriprajna.com/whitepapers/architecture-trust-synthetic-deception-arup-deepfake-breach