When Lightning Strikes Thrice: Breaking Thunderbolt 3 Security

Thunderbolt is a high-bandwidth interconnect promoted by Intel and included in laptops, desktops, and other systems. Being PCIe-based, Thunderbolt devices possess Direct Memory Access (DMA)-enabled I/O. In an "evil maid" DMA attack, where adversaries obtain brief physical access to the victim system, Maartmann-Moe (Inception), Frisk (PCILeech) and others have shown Thunderbolt to be a viable entry point in stealing data from encrypted drives and reading and writing all of system memory. In response, Intel introduced "Security Levels", a security architecture designed to enable users to authorize trusted Thunderbolt devices only. To further strengthen device authentication, the system is said to provide "cryptographic authentication of connections" to prevent devices from spoofing user-authorized devices. We present Thunderspy, a series of attacks that break all primary security claims for Thunderbolt 1, 2 and 3. So far, our research has found seven vulnerabilities: inadequate firmware verification schemes, weak device authentication scheme, use of unauthenticated device metadata, downgrade attack using backwards compatibility, use of unauthenticated controller configurations, SPI flash interface deficiencies, and no Thunderbolt security on Boot Camp. Finally, we present nine practical exploitation scenarios. In an "evil maid" threat model and varying Security Levels, we demonstrate the ability to create arbitrary Thunderbolt device identities, clone user-authorized Thunderbolt devices, and finally obtain PCIe connectivity to perform DMA attacks. In addition, we show unauthenticated overriding of Security Level configurations, including the ability to disable Thunderbolt security entirely, and restoring Thunderbolt connectivity if the system is restricted to exclusively passing through USB and/or DisplayPort. We conclude with demonstrating the ability to permanently disable Thunderbolt security and block all future firmware updates. All Thunderbolt-equipped systems shipped between 2011-2020 are vulnerable. Some systems providing Kernel DMA Protection, shipping since 2019, are partially vulnerable. The Thunderspy vulnerabilities cannot be fixed in software, impact recently introduced standards such as USB 4 and Thunderbolt 4, and will require a silicon redesign. Finally, we conclude our talk demonstrating our on-going research on designing and implementing protections against Thunderspy.