Borehole · Project · free tier
unslothai/unsloth
Free reading: not for reliance
7f32241 · 11,953 commits · 9,151 files
Read 2026-10-23

unslothai/unsloth: technical due diligence, read from its repository

What Borehole read in unslothai/unsloth's code and history at 7f32241: its licence, security advisories, releases, review practice, tests, dependencies and operations. Every finding cites the file and line it rests on. The full report also covers the team and asks the questions the code cannot answer.

DimensionRating
Architecture & Codebase — including AI provenanceadequate
Engineering Process & SDLC Maturityadequate
Product & Engineering Maturitystrong
Security & Reliability Postureadequate
Observability & Operationsstrong
Fit with the Acquirernot assessable

What the repository says

  • LICENSE: Relicensed: Apache-2.0 → AGPL-3.0 (2026-03-12)

Architecture & Codebase — including AI provenance

adequate
01.6

studio/backend/core/inference/diffusion.py is 14,543 lines, 85× this codebase's median file

adequate

The median source file here is 170 lines. 159 files exceed 1,700 lines, the largest being 14,543. Size is not a defect on its own — some problems genuinely live in one place — but these are where merge conflicts, review fatigue and single-owner knowledge concentrate, and they are the first thing that slows an inheriting engineer down.

Evidence
studio/backend/core/inference/diffusion.py:1 — 14,543 lines (85× median)
studio/install_python_stack.py:1 — 13,744 lines (80× median)
studio/backend/core/inference/video.py:1 — 13,662 lines (80× median)
unsloth/import_fixes.py:1 — 12,906 lines (75× median)
studio/install_llama_prebuilt.py:1 — 12,567 lines (73× median)
01.15

1,843 errors caught and discarded in product code

adequate

1,843 places catch an error and do nothing with it, across 432 files (1.0 per thousand lines). Each is a failure the system will not report. They surface later as problems nobody can explain, and they are cheap to fix now.

Evidence
docker/jupyter/unsloth_branding.py:225 — error caught and discarded
docker/jupyter/unsloth_branding.py:230 — error caught and discarded
docker/unsloth_nb_strip_colab.py:51 — error caught and discarded
01.25

2 base images float: no version, or `latest`

adequate

2 of 4 `FROM` lines in product Dockerfiles name an image without a version, or with `latest`. Each build pulls whatever the registry serves that day, so the image that ships can change without a commit, and last month's build cannot be reproduced. Pin a version, or a digest.

Evidence
docker/Dockerfile:368 — base image without a version
docker/Dockerfile:397 — base image without a version

Engineering Process & SDLC Maturity

adequate
02.24

5% of recent commits change more than 2,000 lines

adequate

512 of 10119 commits in the last 24 months changed over 2,000 lines of code, not counting reformatting, renames, vendored or generated code. A change that size is rarely reviewed line by line. Ask how they were reviewed, and what made them that large.

Evidence
512 commits over 2,000 lines

Product & Engineering Maturity

strong

No finding.

Security & Reliability Posture

adequate
05.1

1 credential-shaped string committed to the tree

adequate

Matches for JWT appear in tracked files outside test and fixture paths. They stay in the history even after they leave the tree, so what matters is whether each was rotated. The code cannot tell whether any of them is live, and that decides whether this matters: none sits in production configuration, and none is in a format its provider keeps for live keys. Is any of these live? Show that each was rotated after it was committed, and when.

Evidence
JWT pattern matched
05.6

1 Dockerfile runs the application as root

adequate

No USER instruction switches away from root, so the process runs with full privileges inside the container. It is a one-line fix and an ordinary expectation, which is why its absence is worth asking about: it usually says more about whether anyone has reviewed the deployment than about this container specifically.

Evidence
docker/Dockerfile — runs as root
05.27

The build runs third-party code it does not pin

adequate

24 steps pipe a download straight into a shell. Tags have been moved to malicious commits in real supply-chain attacks, and a piped script is whatever the server returns that minute. Pinning to a commit is the fix, and it is small.

Evidence
.github/scripts/agent-guides-install.sh:63 — download piped into a shell
05.20

27 SQL statements built from strings

adequate

27 places in product code build SQL by formatting or joining strings rather than passing parameters, and in 13 of them a value, not a table or column name, goes into the string. Where any of those values comes from a request, it is an injection point. Each site needs a look; the fix, parameters, is small.

Evidence
studio/backend/auth/storage.py:597 — SQL built from a string
studio/backend/auth/storage.py:601 — SQL built from a string
studio/backend/core/rag/store.py:577 — SQL built from a string
05.21

14 shell commands built from strings

adequate

14 places run a command through a shell (`shell=True`, `os.system`, `exec` with a template) rather than passing arguments. Where any part comes from outside, it is command injection. Ask where each one's input comes from.

Evidence
studio/backend/core/inference/tool_path_approval.py:2426 — command run through a shell
studio/backend/core/inference/tool_path_approval.py:3184 — command run through a shell
unsloth/_gpu_init.py:714 — command run through a shell
05.22

2 deserialisers that can run code

adequate

2 places read data with a deserialiser that executes what it reads (pickle, `yaml.load` without a safe loader). That is safe on data the system wrote itself and remote code execution on anything else. Ask where each one's input comes from.

Evidence
studio/backend/core/inference/tool_path_approval.py:3407 — code-executing deserialiser
unsloth_cli/commands/eval.py:309 — code-executing deserialiser
05.26

15 locked dependencies have published vulnerabilities, 5 high or critical in what ships

adequate

15 of 1,764 locked dependencies match 26 published advisories in OSV, 6 of them rated high or critical. A lockfile says which version is installed, not whether the flawed code is reachable from this product, so each needs a look. Most are fixed by an upgrade, and the advisory names the version. This lockfile pins development tools alongside what ships and does not say which is which, so some of these may never reach a user.

Evidence
studio/frontend/package-lock.json — dompurify 3.4.16: GHSA-cg2g-xh64-v4q7 (critical), fixed in 3.4.17
studio/frontend/package-lock.json — @modelcontextprotocol/sdk 1.31.0: GHSA-22jm-h49p-29qw (high), fixed in 1.32.0
studio/frontend/package-lock.json — node-forge 1.4.0: GHSA-86w9-cpqp-85rv (high)
05.48

6 places insert HTML from a variable without sanitising it

adequate

6 places set HTML from a value that is neither a literal nor passed through a sanitiser. Where the value can come from a user, that is stored cross-site scripting, and it is the first thing a buyer's security review looks for in a frontend.

05.8

Manifest says Apache-2.0, licence file says AGPL-3.0

adequate

pyproject.toml declares Apache-2.0. COPYING contains the text of AGPL-3.0. Package managers and downstream users read the first; auditors and acquirers read the second. One of them is wrong, and until it is settled the terms this code is offered under are genuinely ambiguous.

Evidence
pyproject.toml:16 — declares Apache-2.0
COPYING — text of AGPL-3.0
05.60

Relicensed: Apache-2.0 → AGPL-3.0 (2026-03-12)

adequate

The root licence changed 1 time(s). The latest terms restrict what users, hosts or a buyer may do in ways the earlier ones did not. Code taken under the old terms stays under them, so both matter in diligence.

Evidence
LICENSE — Apache-2.0 from 2023-11-29
COPYING — AGPL-3.0 from 2026-03-12
05.65

2 advisories against this project in the last 24 months, 1 high or critical

adequate

The newest was published on 2026-10-19. A project that publishes advisories is telling its users about its fixes, and the count and pace say how much security work it carries. None of them affects the surveyed version.

Evidence
2026-10-19: CVE-2026-112075 (low): A weakness has been identified in unslothai unsloth up to 0.1.501-beta. This affects the function FastSentenceTransformer._is_transformer_module_ref of the file
2026-09-28: CVE-2026-93348 (high): Unsloth Zoo versions 2025.9.9 before 2026.8.14, as implemented in Unsloth 2025.9.9 through 2026.8.19, contains a code injection vulnerability in the model-loadi

Observability & Operations

strong
06.9

No error reporter is wired in

adequate

Nothing in the service initialises an error reporter (Sentry, Rollbar, Bugsnag, Honeybadger or an APM agent). An exception in production reaches a person only if someone is reading the logs, which usually means a customer reports it first.

Evidence
no error reporter initialised

Fit with the Acquirer

not assessable

Not rated: structurally not assessable from a repository.

Read more, or read your own

The full report on unslothai/unsloth: every dimension, the questions for the room, and the audit log.

Survey a repository: free on public repositories. Private code is read on your own machine by a collector whose source you can read first.