Physical Attacks Bypass Intel SGX and AMD SEV-SNP Trusted Enclave Protections

Researchers from KU Leuven and the University of Birmingham have demonstrated new physical attacks that undermine the security guarantees of trusted execution environments (TEEs) provided by Intel and AMD processors. These TEEs, specifically Intel's Software Guard Extensions (SGX) and AMD's Secure Encrypted Virtualization with Secure Nested Paging (SEV-SNP), are widely used in cloud computing to protect sensitive data and operations from attackers, even those with administrative access to the underlying infrastructure. The attacks, named Battering RAM and Wiretap, exploit vulnerabilities in the deterministic encryption schemes used by both SGX and SEV-SNP. By inserting a low-cost hardware interposer—costing as little as $50—between the CPU and DDR4 memory modules, attackers can observe and manipulate encrypted data in real time. The Battering RAM attack allows adversaries to not only view but also modify encrypted data, enabling the introduction of software backdoors or data corruption within protected enclaves. The Wiretap attack, on the other hand, enables passive decryption of sensitive data without detection. These attacks are particularly concerning for cloud environments, where confidential computing workloads rely on hardware-level memory encryption to secure data from potentially untrusted cloud service providers. The interposer operates transparently during system startup, passing all trust checks, and can later be activated to redirect protected memory addresses to attacker-controlled locations. On Intel platforms, Battering RAM enables arbitrary read and write access to enclave memory, while on AMD systems, it bypasses recent firmware mitigations and allows the introduction of arbitrary backdoors into virtual machines. The attacks highlight fundamental limitations in the current design of hardware-based memory encryption, especially when deterministic encryption is used. All systems using DDR4 memory and relying on these TEEs for confidential computing are potentially vulnerable. The research underscores the need for more robust hardware security mechanisms and may prompt cloud providers and customers to reassess their trust in existing enclave technologies. The findings have significant implications for the security of cloud-hosted applications, including those used by messaging platforms like Signal and WhatsApp, which depend on these protections to safeguard user data. The attacks demonstrate that physical access to cloud hardware, even with minimal resources, can compromise the confidentiality and integrity of sensitive workloads. Both Intel and AMD are expected to review these findings and consider architectural changes or additional mitigations in future hardware generations. The research also raises questions about the adequacy of current cloud security recommendations and the potential for insider threats within data centers. Organizations relying on SGX or SEV-SNP for critical workloads should closely monitor vendor advisories and consider additional layers of defense. The academic community continues to play a vital role in uncovering and publicizing such vulnerabilities, driving improvements in hardware security.