Power analysis attacks are not new, but they are typically local and require some sort of power measurement interface in order to monitor differences between computations and infer details about what’s being processed. This side-channel technique has been used before to leak cryptographic secrets such as keys.
Hertzbleed, however, shows that frequency scaling generates timing differences in computations and these can be observed even remotely without any power measurement interface. The novelty is that Hertzbleed works even against so-called constant time cryptographic implementations that were intentionally designed to prevent leaking information through timing analysis.
The researchers used Hertzbleed to implement a novel chosen-ciphertext attack against SIKE (Supersingular Isogeny Key Encapsulation), a post-quantum key encapsulation mechanism that is also a NIST competition finalist and is implemented as constant time. The team was able to perform a full key extraction via remote timing.
Intel published guidance for developers of cryptographic libraries to mitigate Hertzbleed using software countermeasures. Another possible mitigation is to disable “Turbo Boost” at runtime on the system, but this has a significant system-wide performance impact.
SQUIP (CVE-2021-46778)
SQUIP is a side channel attack and vulnerability impacting AMD CPUs that was disclosed in August 2022. The attack was devised by researchers from Lamarr Security Research, Graz University of Technology, and Georgia Institute of Technology, and it exploits scheduler queues used during simultaneous multithreading (SMT) operations to schedule instructions that will be executed in CPUs. By measuring the contention level on scheduler queues an attacker may potentially leak sensitive information, AMD said.
Zenbleed (CVE-2023-20593)
Zenbleed is a vulnerability patched in July 2013 in the Zen 2 family of AMD CPUs. The flaw was found by security researchers from Google and is described as a user-after-free memory vulnerability but for CPUs. It’s caused by incorrectly implemented speculative execution of the SIMD Zeroupper instruction and can allow attackers to leak stale data from physical hardware registers. Such data can include sensitive information such as passwords or encryption keys.
Downfall (CVE-2022-40982)
Downfall, technically called Gather Data Sampling (GDS) by Intel, is a transient execution vulnerability disclosed in August 2023 that impacts multiple generations of Intel CPUs. Found by security researchers from Google, the flaw is similar to Zenbleed in that it allows attackers to leak sensitive data belonging to other processes and users sharing the same CPU core because stale data stored in physical hardware registers as a result of speculative execution is forwarded to subsequent instructions. The data can be extracted using techniques similar to those used by Meltdown. The flaw also impacts the security of Intel’s Software Guard Extensions (SGX) security subsystem.
Reptar (CVE-2023-23583)
Reptar is a third CPU vulnerability found by Google security researchers last year and was patched in November 2023. It impacts Intel CPUs that support a new feature called fast short repeat move (FSRM) and can result in privilege escalation. The flaw is caused by the CPU microcode not ignoring redundant instruction prefixes when FSRM is active and interpreting them in weird ways.
Inception (CVE-2023-20569)
Inception is a vulnerability in AMD CPUs that can lead to found by researchers from ETH Zurich that was disclosed in August 2023 and can lead to sensitive information disclosure. Inception is a new type of speculative execution attack that hijacks the transient control-flow of return instructions and allows attackers to insert new predictions into the CPU branch predictor at an attacker-controlled address register.
SLAM
Spectre based on Linear Address Masking (SLAM) is a proof-of-concept attack technique devised by researchers from Vrije Universiteit Amsterdam that shows how previously unexplored Spectre gadgets could be exploited on upcoming AMD, Intel, and ARM CPUs that implement linear address masking, a new security feature planned by CPU vendors: Intel’s Linear Address Masking (LAM), AMD’s Upper Address Ignore (UAI), and ARM’s Top Byte Ignore (TBI). SLAM is notable for being the first speculative execution attack targeting CPU features that were announced but not yet released.
GhostRace (CVE-2024-2193)
GhostRace is a new type of CPU attack disclosed in March 2024 by researchers from Vrije Universiteit Amsterdam that take advantage of race conditions on speculatively executed code paths. The research shows that synchronization primitives implemented using conditional branches at the OS level can be bypassed on speculative paths using a Spectre v1 attack, potentially allowing for information leaks from targeted software.
TikTag
TikTag is an attack that leverages speculative execution to bypass a new security feature in ARM CPUs called the Arm Memory Tagging Extension (MTE). This feature, when used by operating systems, makes it harder to exploit out-of-bounds memory violations such as buffer overflows that can lead to arbitrary code execution. The TikTag attack was developed by a team of researchers from Seoul National University, Samsung Research and Georgia Institute of Technology and was described in a research paper in June 2024. Separately, researchers from Vrije Universiteit Amsterdam already showed that MTE is vulnerable to speculative execution probing with an attack they dubbed Spectre-MTE and proposed a proposed a mitigation called StickyTags.
Indirector
Indirector is a new speculative execution attack that is a variation of Spectre v2 and was disclosed in July 2024. The attack, developed by researchers from University of California San Diego exploits the indirect branch predictor (IBP) and the branch target buffer (BTB) in high-end Intel CPUs (Raptor Lake and Alder Lake) to perform precise branch target injections and leak sensitive data across processes and privilege levels.
Register File Data Sampling (CVE-2023-28746)
Disclosed in March 2024, Register File Data Sampling (RFDS) affects some Intel Atom processors and allows local software to infer stale data left in floating-point, vector, or integer registers after transient execution. RFDS does not let an attacker choose which values are leaked and exposes only the data that happens to remain in the register files. Intel released a microcode update.
SinkClose (CVE-2023-31315)
Disclosed in August 2024 by researchers from IOActive, SinkClose gives applications with kernel-level (ring 0) access to modify the configuration of System Management Mode (SMM) even when the SMI lock is enabled. The flaw went undetected for almost two decades and affects nearly all AMD EPYC, Ryzen, and Threadripper CPUs. It provides privilege escalation to ring -2, the processor’s most privileged execution mode, which enables the deployment of malicious implants into the UEFI firmware. AMD released microcode updates for some of the affected chips, but their availability depends on motherboard vendors providing UEFI updates.
Native BHI (CVE-2024-2201)
In April 2024 researchers from Vrije Universiteit Amsterdam demonstrated that the Branch History Injection (BHI) Spectre v2 variant can be exploited without unprivileged eBPF. The researchers built a tool called InSpectre Gadget that used symbolic execution to find usable Spectre gadgets and identified more than 1,500 in the Linux kernel. Using those gadgets, they built the first native BHI exploit, which leaked kernel memory at 3.5KB/s on Intel CPUs. Intel’s BHI_DIS_S mitigation and software defenses are the available countermeasures.
GoFetch
GoFetch is a side-channel attack that exploits a data memory-dependent prefetcher (DMP) in Apple’s M-series CPUs. Disclosed in March 2024 by a team of academic researchers from different universities, the technique was used to extract keys from constant-time implementations of OpenSSL’s Diffie-Hellman key exchange, Go’s RSA decryption, and the post-quantum CRYSTALS-Kyber and CRYSTALS-Dilithium algorithms. DMP can be disabled on M3 CPUs through the data-independent timing (DIT) bit, and a hidden configuration bit can disable it on M1 and M2.
EntrySign (CVE-2024-36347 and CVE-2024-56161)
Two flaws disclosed in March 2025 by researchers from Google and dubbed EntrySign can be used to forge the signature of microcode patches for AMD Zen CPUs. EntrySign affects Zen 1 through Zen 5 CPUs and can undermine AMD’s Secure Encrypted Virtualization-Secure Nested Paging (SEV-SNP) because it lets attackers with local administrator privileges load malicious microcode into confidential guests. AMD released microcode and Linux kernel changes.
Transient Scheduler Attacks (CVE-2024-36348, CVE-2024-36349, CVE-2024-36350, and CVE-2024-36357)
Transient Scheduler Attacks (TSA) are two timing side-channel vulnerabilities in Zen 3 and Zen 4 CPUs that were found by AMD while investigating a Microsoft research report. The flaws allow attackers to leak data across privilege boundaries from the store queue and the L1 data cache. AMD rated the two most serious variants medium severity and released Platform Initialization firmware and operating system updates.
Branch Privilege Injection (CVE-2024-45332)
The hardware mitigations built into Intel CPUs to stop Spectre v2 branch target injection attacks contain a race condition reported in May 2025 by researchers from ETH Zurich. The flaw affects Intel CPUs released since 2018, from Coffee Lake Refresh through Raptor Lake, and was used to leak arbitrary kernel memory on a fully patched Ubuntu 24.04 system with default mitigations enabled. Intel released a microcode update.
VMScape (CVE-2025-40300)
Disclosed in September 2025 by researchers from ETH Zurich, VMScape is a Spectre-like branch target injection attack that breaks the isolation between guest virtual machines and the host hypervisor on AMD and Intel CPUs. The researchers used the attack to extract the disk encryption key from a KVM/QEMU host. Linux kernel maintainers mitigated the issue by adding an indirect branch prediction barrier on each VM exit, with only marginal performance overhead.
SysBumps (CVE-2024-54531)
SysBumps is the first attack to break kernel address space layout randomization (KASLR) on macOS running on Apple M-series chips with kernel isolation enabled. It was described in a research paper published in December 2024 by researchers from Soongsil University and is a Spectre-like vulnerability that occurs when macOS handles user-supplied pointers. This allows unprivileged attackers to force the processor to speculatively translate attacker-chosen kernel addresses, reveling whether an address is valid.
LoongLeak
LoongLeak is an architectural vulnerability in Loongson 3A5000 and 3A6000 processors, which use China’s LoongArch instruction set, that was described in a research paper presented at USENIX Security 2026 by researchers from Germany’s CISPA Helmholtz Center for Information Security. It belongs to the same class of architectural flaws as Reptar and Zenbleed, which leak data from the retired instruction stream rather than only during transient execution. The flaw can be used to recover kernel AES disk-encryption keys, partial root password hashes, and break address space layout randomization (ASLR) within seconds. The software workaround involves emulating floating-point instructions in the kernel but carries a 10x to 21x performance penalty.
DRAM and GPU memory attacks
- Rowhammer
- Rowhammer.js
- Drammer
- Flip Feng Shui
- ECCploit
- Throwhammer
- RAMBleed
- ZenHammer
- Phoenix
- LeftoverLocals
- GPUHammer
- GPUBreach
- GPUThor
- BadRAM
- Battering RAM
- DDRop
Rowhammer
Rowhammer is a physical effect with security implications that occurs inside SDRAM chips when the same physical row of memory cells is read for a large number of times in rapid succession — an action dubbed hammering. This can cause electric charges from cells in the hammered row to leak into adjacent rows, modifying the value of the cells in those rows. This is known as bit flipping and possible because of the increased cell density of modern SDRAM chips, particularly DDR3 and DDR4.
While the Rowhammer effect has been known or documented for a long time, members of Google’s Project Zero team were the first to prove it can have security implications in March 2015 when they revealed two privilege escalation exploits based on it.
Rowhammer.js
Rowhammer.js was an implementation of the Rowhammer attack via JavaScript, proving that this flaw can be exploited remotely through the browser, simply by visiting a malicious web page. Browser vendors have added mitigations against this exploit.
Drammer – CVE-2016-6728
Drammer is a Rowhammer-type exploit demonstrated in 2016 against Android devices. Until then the memory chips in mobile devices were thought to be unaffected.
Flip Feng Shui
An implementation of the Rowhammer attack against virtual machines, where a malicious guest VM can flip bits in the physical memory affecting a different virtual machine in a controlled manner. The researchers demonstrated this by breaking the OpenSSH public key authentication in the target VM.
ECCploit
ECCploit is an attack that demonstrates that Rowhammer-type attacks can work even against SDRAM chips that have error-correcting code (ECC) capabilities. This type of memory, which is typically used in servers, was thought to be immune to Rowhammer.
Throwhammer
A Rowhammer attack that can be exploited over a network by leveraging the remote direct memory access (RDMA) feature present in fast network cards like those used in servers.
RAMBleed
RAMBleed is the first attack that has shown it is possible to use the Rowhammer effect to steal data from memory cells instead of simply modifying it. Previous Rowhammer attacks compromised memory integrity through bit flips, which could lead to privilege escalation and other conditions. Meanwhile, RAMBleed uses row hammering and a side-channel in order to infer information about and ultimately extract data from adjacent memory cells. In that respect it is similar to the effects of Meltdown and Spectre.
ZenHammer
Disclosed in March 2024 by researchers from ETH Zurich, ZenHammer was the first Rowhammer attack on AMD Zen-based CPUs and also the first to trigger bit flips in DDR5 memory. The researchers reverse-engineered AMD’s undisclosed DRAM addressing functions, synchronized their access patterns with refresh commands to evade Target Row Refresh (TRR) mitigations, and scheduled flush and fence instructions to reach a high activation rate. AMD published a security bulletin and pointed users to their DRAM and system vendors for mitigation.
Phoenix (CVE-2025-6202)
Disclosed in September 2025 by researchers from ETH Zurich and Google, Phoenix is a Rowhammer technique that defeats the advanced TRR mitigations in DDR5 memory and achieves privilege escalation on systems with AMD CPUs. The researchers used the bit flips to modify page table entries, leak RSA-2048 SSH keys from a co-located virtual machine, and gain root through the sudo binary in as little as 109 seconds. The attack works only on AMD platforms because Intel CPUs add a pTRR mitigation.
LeftoverLocals (CVE-2023-4969)
GPUs from Apple, AMD, and Qualcomm can leak data left in local memory by other processes. The flaw was disclosed in January 2024 by researchers from Trail of Bits who managed to recover roughly 5.5MB of data per GPU invocation on an AMD Radeon RX 7900 XT, or about 181MB for each query when a large language model is running on llama.cpp, enough to reconstruct the model’s responses. Apple fixed the issue in its A17 and M3 chips, and fixes were released for some Qualcomm and Imagination devices.
GPUHammer
The first Rowhammer attack against an Nvidia dedicated GPU with GDDR6, GPUHammer was disclosed in August 2025 by researchers from the University of Toronto. The attack showed that the GDDR6 memory used by Nvidia’s RTX A6000 can be made to flip bits from an unprivileged CUDA program. Nvidia recommends enabling error-correcting code (ECC), which costs up to 10% performance and 6.25% memory capacity on the A6000.
GPUBreach
GPUBreach is a Rowhammer attack that corrupts GPU page tables to escalate privileges. It was discovered by researchers from the University of Toronto and builds on GPUHammer. An unprivileged CUDA kernel can use targeted page-table corruption to read and write another process’s GPU memory, leak cryptographic keys from GPU libraries, and tamper with a model’s GPU assembly. The researchers also used memory-safety bugs in Nvidia’s kernel driver to gain kernel write privileges on the CPU and spawn a root shell, even with the input-output memory management unit (IOMMU) enabled.
GPUThor
Disclosed in August 2026 by researchers from the University of Toronto, GPUThor is a Rowhammer attack that defeats the ECC protection on some Nvidia enterprise GPUs. The earlier GPUHammer and GPUBreach attacks used uniform hammering patterns, which the Target Row Refresh (TRR) mechanism in newer memory can detect and ECC can correct. GPUThor enables the first practical non-uniform hammering on GPU DRAM, producing double- and even triple-bit errors that ECC was not designed to handle. Users should avoid running untrusted code on vulnerable GPUs and monitor Nvidia’s ECC telemetry.
BadRAM (CVE-2024-21944)
Tampering with the Serial Presence Detect (SPD) chip on a DRAM module can make it report twice its actual capacity and defeat the memory protections in trusted execution environments such as those based on AMD SEV-SNP. The attack was disclosed in December 2024 by researchers from KU Leuven, the University of Lübeck, and the University of Birmingham. The attack costs about $10 and, on some modules with unlocked SPD chips, can be carried out without physical access. AMD released firmware that validates the memory configuration during boot.
Battering RAM
A runtime memory-aliasing attack that uses a low-cost DDR4 interposer between the processor and a memory module. The attack was presented at Black Hat Europe in December 2025 by researchers from KU Leuven and requires a $50 device to manipulate memory address mapping at runtime, bypassing the boot-time checks Intel and AMD added after BadRAM. The researchers gained arbitrary plaintext read and write access to memory protected by Intel Software Guard Extensions (SGX), extracted the platform provisioning key, forged attestation reports, and implanted backdoors in AMD SEV-protected virtual machines. Neither Intel nor AMD plan a fix because the attack requires physical access and falls outside their threat models.
DDRop
Disclosed in September 2026 by researchers from KU Leuven, ETH Zurich, Durham University, and Google, DDRop is a memory-bus attack that breaks the memory-integrity guarantees of Intel Trusted Domain Extensions (TDX), Intel Scalable SGX, and AMD SEV-SNP. A $159 interposer forces an error on the DDR5 command bus and suppresses the signal the memory module uses to report it. On Intel TDX, dropped writes during page-table setup let the researchers inject their own page-table entries, read protected memory, force a confidential virtual machine into debug mode, and forge remote-attestation reports with attacker-chosen measurements. Intel and AMD classified the technique as outside their published threat models.
Editor’s note: This article, originally published in July 2019 and amended in August 2022, has been updated to include new vulnerabilities as they come to light.
