Exposing Bad Consumer Crypto - AI vs. Encrypted Drives

Episode Summary

Security Now examines how AI is changing vulnerability research and security risks, focusing on consumer USB drives marketed as encrypted that reportedly offered little or no real encryption.

Consumer Crypto and Encrypted Drives

Steve Gibson discusses findings by security researcher Binary Ninja (Shu Lee), who used reverse engineering and AI-assisted analysis to investigate consumer USB security products, including a Lenovo ThinkPlus fingerprint USB drive and several other drives and encryption utilities.

• The Lenovo drive was marketed as AES-256 encrypted, but the researcher found an ordinary SD card containing unencrypted data. The fingerprint reader controlled whether the drive connected to a computer; it did not encrypt or decrypt files.
• Other products had serious design flaws, including fixed passwords or keys that could bypass the user’s password. One utility reportedly embedded its AES key in the executable, logged passwords, used weak or inappropriate cryptographic practices, and offered a “fast mode” that encrypted only part of some files.
• An AES-256 claim alone does not establish security. Proper encryption requires sound key generation and handling, a user secret that genuinely protects the encryption key, and correct use of the algorithm. Steve cautions that fingerprint unlocking is not automatically secure: biometric matching is approximate, so the device must have a key available internally once it accepts a match.
• Binary Ninja says AI helped identify flaws quickly, including reproducing earlier reverse-engineering work in minutes. Steve sees this as a troubling new risk: AI can make opaque, poorly designed crypto products much easier for non-experts to analyze and bypass.

Practical takeaway: Be skeptical of proprietary “encrypted” USB products and bundled encryption software. Favor independently reviewable, open-source encryption tools over unverified manufacturer utilities.

Other Security Developments

• Linux kernel flaw: Steve discusses an AI-discovered kernel race condition involving a use-after-free in networking code. The severity was disputed, and successful exploitation had not been demonstrated. The episode recommends updating Linux systems, while noting that exploitation would require difficult timing and exposure of relevant TCP services.
• PXE/UEFI buffer overflow: A separate, long-standing flaw in EDK2 PXE boot code could allow code execution during network boot. The reported bug involved an incorrect 504-byte copy limit in ICMP handling. An attacker would need network access and an opportunity to target a machine during PXE boot; home users who do not network-boot are unlikely to be affected. Suggested mitigations include disabling PXE boot where it is not needed, updating supported firmware, and restricting ICMP error traffic across sensitive network boundaries.
• Cisco AI refactoring plan: Cisco reportedly assigned a 27-year-old employee to lead an effort to rewrite 98% of its code using AI, with a two-year goal. Steve questions whether such a sweeping rewrite addresses Cisco’s more fundamental security concerns, particularly insecure defaults and policies that can leave products exposed.
• Crypto thefts: The hosts report a $6.3 billion Binance theft, with some funds frozen and the company pledging to cover user losses. They also discuss 710 Polymarket wallets drained of about $2.7 million after a compromise involving the third-party web-monitoring provider DebugBear. Polymarket said it would reimburse affected users and losses related to pauses.
• AI crawler opt-out: Major AI companies were reported to be backing a shared robots.txt signal for sites that do not want their content used for AI training. The signal is voluntary, so sites still depend on crawlers honoring it.
• AI in orbit: Google’s experimental satellite ran a Gemma model on onboard Tensor Processing Units and generated a haiku. The test is part of research into operating AI hardware in space.
• Listener topics: Steve rejects his earlier idea of “stealth” security updates, discusses using Cloudflare tunnels and a web application firewall to keep a server’s inbound ports off the public internet, and argues that a universal AI kill switch is unlikely once models can run locally or on private infrastructure. A small company’s report of an AI-assisted attack illustrates the growing security burden on smaller organizations.

Reviews, Recommendations, and Guides

• Book recommendation: Steve recommends Kevin Roose’s The AGI Chronicles, a readable account of the people and events behind the development of advanced AI.
• Reverse-engineering tools: Binary Ninja’s account describes using time-travel debugging to record and replay program execution, alongside AI-assisted analysis. The researcher points readers to Binary Ninja and its MCP server as ways to try the workflow.
• Security guidance: The episode discusses disabling unnecessary PXE boot, updating firmware, filtering network traffic, and evaluating whether encryption products genuinely derive their protection from the user’s secret.

Main Speakers and Sources

• Steve Gibson — primary security analysis and commentary
• Micah Sargent — co-host and discussion partner
• Leo Laporte — sponsor reads

Sources discussed include Binary Ninja, Fortune, Ars Technica, The Verge, and CNBC.