BootBench is the dataset behind SoK: All You Ever Wanted to Know About Bootloader Security but Were Afraid to Ask, by Connor Glosner and Aravind Machiry (Purdue University), IEEE S&P 2026. It pairs open-source bootloader source trees with the vulnerabilities recorded against them, so that bootloader analysis tools can be evaluated against real bugs in real bootloaders.
Three submodules hold the data. tools/ builds it.
oss-bootloaders — the
corpus. 63 bootloader repositories pinned as nested submodules and grouped by
type (14 Type 1, 28 Type 2, 21 Type 3). Pointers only, nothing vendored. The
SoK identified 43 bootloaders (§ 2.2) and studied 7 in depth (§ 3); the rest
were added afterwards and have no mined data yet.
Two submodules are both named bootloader — type2/bootloader (rust-osdev)
and type3/bootloader (wookey-project). Tooling keyed by directory name sees
62 distinct names, so the second is not scanned and has no wiki page.
bootloader_cve_db —
1,432 CVEs mined from the CVE Project
records and assigned a bootloader type by keyword. Each type directory holds the
full upstream record per CVE, an index (type<N>-results.json), the keyword
breakdown that produced the classification, and generated statistics.
bootloader_vuln_commits —
3,514 commit records mined from those repositories at their pinned revisions:
390 that name a CVE and 3,124 matched on vulnerability keywords. That is down
from 3,656 because re-mining removed 1,151 false positives — 95% of them the
bare dos keyword matching "glados", "TODOs" and "DOS header" — while adding
843 from 17 newly-mined bootloaders. These are the
security fixes that never got a CVE, which is most of them.
cve-commit-links.json
joins the two halves: 78 CVEs resolve to a fixing commit and to the parent
revision that still contains the bug.
analysis-tools/ — 24 bootloader and firmware analysis
tools as submodules, grouped by what they do: static analysis, fuzzing and
rehosting, image inspection, platform assessment. These are what BootBench
exists to evaluate.
oss-bootloaders/DEFENSES.md —
what each bootloader defends itself with: Secure Boot, measured boot, signature
verification, rollback protection, and the mitigations a build actually enables.
30 of the 62 scanned implement Secure Boot; 4 randomise their load address.
wiki/ — a page per bootloader with a timeline figure of its boot,
covering the stages it runs through, how each stage passes state to the next,
what it hands off, and why it is classified as it is; plus a page per analysis
tool with a walkthrough. Start at wiki/Home.md.
tools/ — the collection, classification and validation tools,
one document per tool in tools/docs/.
tools/OVERVIEW.md is the table of which analysis tools
work and which bootloaders each applies to.
scripts/ — the scripts you run: refresh a submodule, add to
the corpus, and scripts/analysis/ to run an analysis tool
against a bootloader. There is no CI in any BootBench repository; these replace
it, and none of them commits or pushes.
PAPERS.md — bootloader papers from the eight venues the SoK
surveyed, grouped by contribution.
Everything is organised by the paper's three-way split, by how a bootloader initialises and hands off:
Type 1, firmware bootloader. Boots from hardware and presents a hardware-agnostic interface to whatever runs next. UEFI/EDK-II, SeaBIOS, coreboot.
Type 2, OS bootloader. Boots from an already-initialised system and prepares for an OS. GRUB2, shim, Windows Boot Manager.
Type 3, monolithic bootloader. Both jobs at once — hardware straight to OS, no handoff. Hardware-specific, common in embedded systems. U-Boot, MCUboot, ARM Trusted Firmware.
Types 1 and 2 chained together are staged booting, typical of desktops, servers and phones. Type 3 is monolithic booting, typical of IoT and MCU devices. The paper defines six attack surfaces over these types: invasive hardware, external hardware, remote access, persistent data sources, post-boot features and boot-time features.
git clone https://github.com/BreakingBoot/BootBench.git
cd BootBench
git submodule update --init oss-bootloaders bootloader_cve_db bootloader_vuln_commits
./scripts/run-all-tools.shThe three data submodules are small. Two other groups are large and optional:
git -C oss-bootloaders submodule update --init --recursive # bootloader sources
git submodule update --init --recursive analysis-tools # the analysis tools
git -C oss-bootloaders submodule update --init type1/edk2 # or just oneYou need the bootloader sources only to re-mine commit histories or to run an analysis tool against source.
To run an analysis tool, use scripts/analysis/run-tool.sh. It works
through docker, so the host needs no toolchain:
./scripts/analysis/run-tool.sh --list # what runs, what is blocked, why
./scripts/analysis/run-tool.sh codeql kexec-tools
./scripts/analysis/run-tool.sh angr /boot/efi/EFI/ubuntu/shimx64.efiOf the 24 tools, 18 have runners verified against real targets. Seven of those
needed a patch, a template or a specific container environment to run at all --
scripts/analysis/PATCHES.md documents each, and
tools/OVERVIEW.md is the table of what works on which
bootloader. The rest need a host database, specific silicon, or a commercial
licence. That spread is the SoK's finding made concrete.
To evaluate a bug-finding tool, pick a fix commit from
bootloader_vuln_commits and check out its parent. That parent is the last
revision that still contains the bug, so it is the version to point a tool at;
the fix commit itself is the ground truth for what the tool should find. Not
every record is a usable target — many describe code paths since removed — so
expect to select rather than run the whole set. The paper does the same, using
seven bootloaders and one commit each.
To study a vulnerability class, start from bootloader_cve_db. Each entry
carries structured cwe_ids and cvss alongside the free-text vuln_type, and
a bootloaders field naming the corpus projects it affects — so "every GRUB
CVE, CWE-787 only, high severity" is a filter rather than a reading exercise.
stats.md breaks each type down by CWE, severity, attack vector and year.
To study defenses, start from oss-bootloaders/DEFENSES.md. It records what
each bootloader implements and, for built artifacts, what its binary actually
enables — the two are different claims and are kept apart.
To scope by attack surface or bootloader type, use the type directories. Type 1 is where the SMM and UEFI firmware bugs are, Type 2 has the Secure Boot bypasses, Type 3 the embedded and MCU work.
Before trusting a number, run validate_dataset.py. It currently passes
every check. It exists because the dataset previously did not: withdrawn CVEs
were still counted, two CVEs were counted under two types, statistics rows did
not sum to their own totals, and previous_commit pointed at the next newer
commit rather than the parent.
To go from a CVE to a vulnerable tree, use
bootloader_vuln_commits/cve-commit-links.json. It resolves 78 CVEs to their
fixing commit and to that commit's parent — the revision to check out to
reproduce the bug. LINKS.md
explains the coverage and why most CVEs have no linked fix.
There is no CI. Each repository has an update script that stages a refresh outside the submodule, prints a diff, and asks before writing:
./scripts/update-bootloader-cve-db.sh # clones cvelistV5 if needed
./scripts/update-bootloader-vuln-commits.sh # needs the full corpus checked out
./scripts/update-oss-bootloaders.sh
./scripts/update-papers.shRead the RELABELLED section before answering yes — those are CVEs that changed
bootloader type. Nothing is committed or pushed; each script prints the git
commands to publish what it changed. See scripts/.
@inproceedings{glosner2026bootloader,
title = {SoK: All You Ever Wanted to Know About Bootloader Security
but Were Afraid to Ask},
author = {Glosner, Connor and Machiry, Aravind},
booktitle = {2026 IEEE Symposium on Security and Privacy (S\&P)},
year = {2026},
doi = {10.1109/SP63933.2026.00163}
}