Piktogramm mit Server auf violettem Hintergrund

Hardware Vulnerabilities

The State of IT Security in Germany in 2026

Hardware vulnerabilities can have serious consequences, because a widespread replacement of affected products is often the only way to remedy them, as the example RAIN shows. Here, the Foreshadow attack on Intel CPUs from 2018 could be carried out successfully even on already patched processors. The same study also shows the effectiveness of additional hardening measures: at a large cloud provider, the attack failed due to supplementary protection mechanisms that remove sensitive memory contents from access from virtual machines. Smartphones are also increasingly in focus, as they increasingly contain dedicated security hardware (cf. SCA_ECDSA). 

The focus of IT security research is also on post-quantum cryptography and the analysis of attacks on its hardware implementations. These algorithms offer protection against attacks by quantum computers. However, implementation of cryptographic algorithms that is secure against attacks on hardware vulnerabilities is a relatively young field of research and therefore challenging. Side-channel attacks, which are not directed at the cryptographic method but at its implementation, such as ML-KEM-SCA, BM, HOCPA, SWOOP, fault attacks (e.g., DILF, XMSS, Rand), or Rowhammer attacks (SLasH-DSA) exploit such implementation weaknesses. 

A concerning trend is the simplification of attack methods with machine learning. Whereas in the past, especially in side-channel attacks, laborious and personnel-intensive manual post-processing of recorded data was necessary, these tasks can now increasingly be automated (e.g., NLM, SnowV). This significantly reduces the effort for attacks and theoretically also enables attackers with fewer resources to carry out such attacks. 

Especially in the environment of cloud computing, there are new attack vectors against established isolation mechanisms. Using the microarchitectural vulnerability VMScape, attackers can compromise the overarching cloud infrastructure from a regular, rented virtual machine and siphon off sensitive data. The new attack method TEE.fail also shows that even modern, hardware-supported confidential computing environments are physically compromisable, in order to obtain sensitive cryptographic keys. 

Another trend is attacks on neural networks exploiting hardware vulnerabilities. Attackers can influence AI models by exploiting microarchitectural vulnerabilities. The model architecture can be reconstructed through latency measurements in SGX enclaves (DNN-Latency-Seq) or through analysis of electromagnetic emissions from GPUs (ModelSpy). Secret model parameters and weights are extracted by means of memory-dependent access patterns in SGX (AF-Harmfull) or by evaluating electromagnetic signals during GPU computations (BarraCUDA). Sensitive user data can be spied on by measuring latencies during power gating of AI accelerators (Gatebleed), electromagnetic emissions (ScaAR), or by analyzing side channels in Mixture-of-Experts LLMs (MoEcho). In addition to information theft, hardware-proximate methods also enable active interference: targeted bit flips in GPU memory (GPUHammer) maliciously modify models and reduce accuracy, while the misuse of hardware mitigations in memory management can enable denial of service (NotSoRefreshing). These attacks are partially possible from a distance of several meters without hardware preparations (ModelSpy, ScaAR).