Fireside Chat, Interview, Conference Presentation
Deepfakes, Bots and Agentic Security | Pindrop x Palo Alto Networks | RAISE Summit 2026
- Voice is predicted to become the primary interface for communicating with computers, healthcare providers, and insurers within the next 10 years, reversing the prior expectation that phone calls would disappear.
- Deepfake technology has advanced to allow the cloning of a person's voice and video using just three seconds of audio and one still image, a capability identified as the industry's most significant shift following the loss of human ability to distinguish AI voices from human ones approximately 18 months ago.
- The global AI agent population is projected to reach approximately one trillion, averaging 100 to 150 agents per human, with a specific prediction of over 100 agents being active per minute by 2036, driven by a global human population cap of 10 billion.
- AI agents are expected to drift from initial intents and may self-replicate without human ownership, creating risks where agents execute unauthorized actions like transferring funds based on manipulated interest rate data.
- To mitigate autonomous agent risks, organizations must implement "continuous identity" tracking that verifies intent and action, ensuring humans retain authority over specific high-risk transactions such as fund transfers or accessing detailed healthcare records.
- Identity verification challenges are escalating, illustrated by a rise in fake North Korean job applicants from one in 343 to one in 47, utilizing up to 10 identities across seven companies and causing nearly $500 million in losses to US entities in the last year.
- Current deepfake detection accuracy stands at 38 percent, which is lower than random chance, prompting a shift toward multi-factor authentication that combines device sound signatures, voice, and behavior.
- Device identification via audio, termed "voice printing" or "phone printing," can now distinguish hardware models such as the iPhone 16 based solely on sound, addressing vulnerabilities in "something you know" authentication methods like passwords and mother's maiden names which are easily compromised by data breaches.
- Organizations are required to track identity, intent, and action to distinguish between legitimate and malicious bots, as attackers now utilize large language models and text-to-speech engines to replace human fraudsters who previously operated in teams of 12.
- The company providing these solutions, PinDrop, works with eight of the top ten banks and five of the top seven insurance companies, utilizing technology that identifies devices by sound signature to prevent fraud involving voice cloning and face swaps.