Authors:	Sandro Rüegge, Johannes Wikner, Kaveh Razavi
Organization:	ETH Zürich
Published at:	34th USENIX Security Symposium
Webpage:	https://comsec.ethz.ch/bprc
Paper:	https://comsec.ethz.ch/wp-content/files/bprc_sec25.pdf
BlackHat USA 25:	https://www.blackhat.com/us-25/briefings/schedule/#-for-privilege-leaking-privileged-memory-from--using-under-embargo-45260

Introduction

We introduce Branch Predictor Race Conditions (BPRC), an event-misordering effect of asynchronous branch predictors. Our work demonstrates multiple variants of BPRC that violate security assumptions on Intel processors, enabling cross-privilege and even cross barrier BTI attacks. This repository contains the paper artifacts to reproduce all presented results.

Requirements

OS: Ubuntu (should run on 20.04, 22.04 and 24.04)

Microarchitectures

Vendor	CPU	Codename	Code	Microarchitecture	experiment_core	experiment_march
Intel	Core i7-14700K	Raptor Lake Refresh	RPL-R	Raptor Cove	2	MARCH_GOLDEN
Intel	Core i7-14700K	Raptor Lake Refresh	RPL-R	Gracemont	16	MARCH_GRACE
Intel	Xeon Silver 4510	Sapphire Rapids	SPR	Golden Cove	2	MARCH_GOLDEN
Intel	Core i7-13700K	Raptor Lake	RPL	Raptor Cove	2	MARCH_GOLDEN
Intel	Core i7-13700K	Raptor Lake	RPL	Gracemont	16	MARCH_GRACE
Intel	Core i7-12700K	Alder Lake	ADL	Golden Cove	2	MARCH_GOLDEN
Intel	Core i7-12700K	Alder Lake	ADL	Gracemont	16	MARCH_GRACE
Intel	Core i7-11700K	Rocket Lake	RKL	Cypress Cove	2	MARCH_CYPRESS
Intel	Core i7-10700K	Comet Lake	CML	Skylake	2	MARCH_SKYLAKE
Intel	Core i9-9900K	Coffee Lake Refresh	CFL-R	Skylake	2	MARCH_SKYLAKE
AMD	Ryzen 9 9900X	Granite Ridge	-	Zen 5	2	MARCH_ZEN5
AMD	Ryzen 7 7700X	Raphael	-	Zen 4	2	MARCH_ZEN4
ARM	Google Tensor	Whitechapel	-	Cortex-X1	4	MARCH_ARMv8
ARM	Google Tensor	Whitechapel	-	Cortex-A76	6	MARCH_ARMv8

Your CPU might have different numbers of cores and thus the Gracemont cores might be at a different index. The experiment_core is used with taskset -c <experiment_core> to determine the core to run on.

Dependencies

Ansible

The experiments were originally run using ansible and the manual execution descriptions are provided for convenience. To run the experiments with ansible you can install ansible as follows.

sudo apt install python3-venv

python3 -m venv .venv
source .venv/bin/activate

pip install ansible

If you already have an ansible inventory you can use the host names from your inventory. Otherwise you can use a hostname or IP address where the READMEs use <hostname> and add an argument with -i <hostname>, (note the comma). Add parameter -k if you need a password for ssh access, add -K parameter if you need a password for sudo.

ansible-playbook -i example.com, run.yaml -e host=example.com

Analysis

Our experiments provide python code to analyze the results. Some of the scripts require you to have matplotlib installed.

sudo apt install python3 python3-pip
pip install matplotlib

x86_64

sudo apt install build-essential git clang linux-headers-$(uname -r) msr-tools libtirpc-dev

armv8

The Android NDK r27c binary tools directory (toolchains/llvm/prebuilt/linux-x86_64/bin/) needs to be available in your $PATH.

Metadata

We store server metadata in the file exp_metadata.py. There is a dummy entry for the manual executions but if you use the ansible scripts on your own hosts you will need to add your servers to the SERVER_MAP in this file.

The key in the map is the hostname of your server (you can run hostname to get it). Each value in the map has at least the code_name and cores attributes. code_name is the code name string for the given processor. cores lists all cores present on the given processor. Each entry is a dict with at least the march attribute which specifies the name of the microarchitecture for the given core. The order of the cores in the list should match the id Linux assigns each core (e.g., when using taskset -c <core-id>).

Overview

Each experiment has their own README.md describing how to run the experiment and what the expected outcomes are.

Folder Structure

bprc
|- ansible                     # Ansible tools to run experiments consistently
\- experiments                 # Experiments from the paper
   |- exp-benchmark-mitigations    # Mitigation performance evaluation
   |- exp-bhi-dis-s                # Intel's BHI_DIS_S behavior
   |- exp-btb-delay                # Branch predictor insertion delays on x86
   |- exp-demo-blackhat            # Attack demo used at BlackHat USA 2025
   |- exp-end2end                  # End-to-end attack
   |- exp-ibp-insertion            # IBP insertion behavior
   |- exp-leak-hypervisor          # Hypervisor vulnerability test
   |- exp-leak-ibpb                # IBPB vulnerability test
   |- exp-leak-rounds              # BPI reliability
   |- exp-leak-supervisor          # Supervisor vulnerability test
   |- exp-leak-supervisor-discern  # Supervisor vulnerability predictor attribution
   |- exp-syscall-split            # BPI syscall delay
   \- uarch-research-fw            # framework with general utils

Experiments by paper section

ID	Section	Reference	Folder Name
-	A.3.2	Artifact Appendix Basic Test	exp-test-setup
E1	Section 5.1	Figure 2	exp-btb-delay
E2	Section 5.2	Table 2, $BPRC_{U \rightarrow K}$	exp-leak-supervisor
E3	Section 5.3	$BPRC_{G \rightarrow H}$	exp-leak-hypervisor
E4	Section 5.3	$BPRC_{IBPB}$	exp-leak-ibpb
E5	Section 6.1	Observation (O4)	exp-ibp-insertion
E6	Section 6.1	Observation (O5)	exp-leak-supervisor-discern
E7	Section 6.2	Figure 8	exp-syscall-split
E8	Section 7.1	Figure 9	exp-leak-rounds
E9	Section 7.1	Table 3	exp-bhi-dis-s
E10	Section 8	Textual	exp-end2end
-	Section 8	Table 4	summary of previous experiments
E11	Section 9	Table 5	exp-benchmark-mitigations

Citation

@inproceedings{bprc,
  title     = {{Branch Privilege Injection}: Compromising Spectre v2 Hardware
               Mitigations by Exploiting Branch Predictor Race Conditions},
  author    = {Sandro Rüegge, Johannes Wikner and Kaveh Razavi},
  booktitle = {34th {USENIX} Security Symposium ({USENIX} Security 25)},
  year      = {2025}
}