Repository navigation
Releases
NHI-Blast 0.3.0 — public beta
NHI-Blast 0.3.0 — public beta
Offline, evidence-labelled analysis of GitHub Actions → AWS OIDC authority: which roles and
credentials are potentially reachable from a hypothetical compromised entrypoint, which
credentials survive revocation (R / E / Z), and a safe rotation order. Local-first Python 3.12+
CLI; no network calls, no execution of scanned content, no credential values stored or printed.
This is a beta pre-release. Verified locally on Windows 11 and by the CI matrix on Ubuntu
and Windows runners (Python 3.12, 3.13 and 3.14); macOS is untested. No independent human
security audit has been performed. Severity labels order findings; they are not verdicts.
Highlights
- Report schema 1.1 (
nhi-blast-report/1.1, additive — 1.0 reports still load): recorded
analysis context (analyzer identity, effective configuration, limits, scenarios, digests) and
explicit completeness for nineteen input families. No score, no "complete" badge. nhi-blast diff BASE HEADandnhi-blast baseline accept|prune: new / changed /
unchanged / resolved / not-observed findings. "Resolved" is claimed only when analyzer,
inputs and completeness are comparable; otherwise the report says "not observed" with reasons.nhi-blast whatif REPORT --plan PLAN: offline containment modelling (disable entrypoint,
remove trust, revoke sessions, rotate) with a Markdown playbook. Refuses to run when the saved
report cannot be replayed exactly. Nothing is executed.- Corrected AWS condition semantics: unknown keys are Unknown, never Absent;
IfExists/Null
no longer produce definite results for keys the analyzer has no information about; numeric,
date, IP, binary and boolean operators evaluated. Results can be broader than 0.2.0 where a
block was wrongly FALSE. - Session lifetime provenance: requested duration bounded by the role maximum and the
one-hour role-chaining limit; the flat 3600 s default is gone. Session policies are recorded,
not evaluated. - New report identity: light and dark themes with a System/Light/Dark switcher, embedded
Host Grotesk + Instrument Sans (SIL OFL 1.1), an Analysis completeness panel, scope-aware
empty states, and Motion micro-interactions with a Motion Off control. Essential values render
independently of decorative animation. - Hardened artifact IO: sensitive-field minimization (kind + length only), staged writes with
rollback/cleanup, actual-open identity tracking, explicit errors for denied output directories,
host paths excluded from serialized metadata.
See the changelog and
release notes
for the full record, including what is not included (diff/what-if HTML, GitHub Action,
GitHub context import, session policy evaluation, public-corpus benchmark, macOS verification).
Downloads
| Asset | Purpose |
|---|---|
nhi_blast-0.3.0-py3-none-any.whl |
Python wheel (Python 3.12+) |
nhi_blast-0.3.0.tar.gz |
Source distribution |
nhi-blast-windows-x64.exe |
Standalone Windows x64 executable (no Python needed; unsigned) |
SHA256SUMS |
SHA-256 digests of the wheel, sdist and executable |
nhi-blast-windows-x64.exe.sha256 |
Standalone checksum record for the executable |
sbom.cdx.json |
CycloneDX 1.6 software bill of materials |
LICENSE, NOTICE, THIRD_PARTY_NOTICES.md |
Licence and embedded-component notices (also embedded in the executable) |
There is no PyPI/TestPyPI publication for this version; install from the assets below.
Install
Python wheel (with the optional HCL extra and pinned runtime dependencies):
pip install "nhi_blast-0.3.0-py3-none-any.whl[hcl]" -c requirements.lock
nhi-blast version
requirements.lock is in the source distribution and the repository at tag v0.3.0.
Windows executable: download nhi-blast-windows-x64.exe and run it directly:
nhi-blast-windows-x64.exe version
nhi-blast-windows-x64.exe scan path\to\repo --simulate-compromise action:owner/action@ref
The executable is unsigned and carries LICENSE, NOTICE and THIRD_PARTY_NOTICES.md embedded.
Verify
sha256sum -c SHA256SUMS
See docs/verify-release.md
for the full verification guide. The wheel and sdist carry GitHub build attestations from the
release workflow. The Windows executable is built locally from the tagged source with the locked
toolchain; it is covered by SHA256SUMS and its own .sha256 record, not by a build attestation.
Quickstart
nhi-blast scan tests/fixtures/fx07-incident \
--simulate-compromise action:tj-actions/changed-files@v45 \
--format json --format sarif --format html --fail-on high
Licence
MIT. Published releases 0.1.0, 0.1.1 and 0.2.0 keep their original terms. Embedded fonts are
SIL OFL 1.1; the vendored Motion runtime is MIT. See THIRD_PARTY_NOTICES.md.
NHI-Blast v0.2.0 — public beta
NHI-Blast v0.2.0 — public beta
This is a pre-release of NHI-Blast 0.2.0, a public beta of the offline blast-radius
analyser for non-human identity in CI. It is provided for testing and evaluation. It is
not a production security control, and its analysis is not a substitute for an
independent cloud security assessment. No independent audit has been performed.
Highlights
- Standalone Windows x64 executable.
nhi-blast-windows-x64.exeis a one-file
console application that bundles Python 3.12, the optional Terraform HCL parser, the
console extras and every report asset. It needs no Python, pip, administrator rights or
source checkout, makes no network requests and requires no cloud credentials. - Redesigned offline HTML report. The report uses the Vermilion Noir visual system,
a deterministic layered SVG authority graph, an R/E/Z survival rail, a rotation plan,
accessible graph navigation and a print/PDF layout. The report remains a single
self-contained file with a restrictive content security policy and no external
requests. - MIT License. The 0.2.0 source tree and this release are licensed under the MIT
License. Versions 0.1.0 and 0.1.1 remain under the MIT License they were published
with.THIRD_PARTY_NOTICES.mdrecords the Python runtime and every bundled package
with the full upstream licence texts.
Install
Windows standalone executable
-
Download
nhi-blast-windows-x64.exeand
nhi-blast-windows-x64.exe.sha256from this release. -
Verify the download in PowerShell before running it:
$record = (Get-Content .\nhi-blast-windows-x64.exe.sha256 -Raw).Trim() $actual = (Get-FileHash .\nhi-blast-windows-x64.exe -Algorithm SHA256).Hash.ToLower() if ($record -ne "$actual nhi-blast-windows-x64.exe") { throw "checksum mismatch" }
-
Run
nhi-blast-windows-x64.exe --help, then
nhi-blast-windows-x64.exe scan . --format html --out nhi-blast-out.
The executable is unsigned. Windows may show a reputation-based SmartScreen warning; no
code-signing certificate is used and no SmartScreen reputation is claimed.
Python source
Requires Python 3.12, 3.13 or 3.14.
python -m venv .venv
.venv/bin/python -m pip install ".[hcl]"
On Windows use .venv\Scripts\python.exe. [hcl] adds the optional Terraform HCL
fallback. Installation from an index is required once; scans run offline.
Supported platforms
- Windows 11 x64 with the standalone executable.
- Python 3.12, 3.13 and 3.14 on Windows and Linux, from source.
- GitHub Actions
ubuntu-latestandwindows-latestrunners in CI.
Other platforms and architectures are untested. The executable is built for Windows x64
only. No package-index publication is made; this is a GitHub pre-release.
Limitations and security warnings
- The tool reads GitHub Actions workflows, Terraform plan/state JSON with an HCL
fallback, and IAM policy JSON. It does not execute scanned content, fetch anything or
evaluate expressions. - Not modelled: SCPs, permission boundaries, KMS key policies, environment protection,
branch rules and live AWS state. Seedocs/known-limitations.mdin the source tree. - Findings describe potential authority and conditional revocation survival. They do not
prove compromise, credential access or role takeover. - Credential values are never stored, printed, logged or embedded. If you believe an
output contains a credential value, treat it as sensitive and report it privately as
described inSECURITY.md. - The beta has no SLA and no fix commitment. Back up your work and validate results
before acting.
Beta testing
Use synthetic fixtures such as tests/fixtures/fx07-incident, or repositories you own or
are explicitly authorised to assess. Do not send credentials, secrets, private repository
contents or live cloud data. Report reproducible problems through the beta bug report
issue form with the operating system, version, command, expected result, actual result
and sanitised logs. The full process is in docs/beta-testing.md.
Licence
MIT License. See LICENSE, NOTICE and THIRD_PARTY_NOTICES.md, which are attached to
this release and embedded in the executable.
NHI-Blast 0.1.1
NHI-Blast 0.1.1
Released on 2026-10-27 as the GitHub Release
v0.1.1 with the wheel,
sdist and SHA256SUMS. No package-index upload, deployment or SaaS was performed.
This is a report-focused patch; the analysis pipeline, CLI commands, JSON schema
(nhi-blast-report/1.0) and SARIF 2.1.0 output are unchanged from 0.1.0. All three
assets were re-downloaded after publication and verified against SHA256SUMS.
What changed
- The single-file offline HTML report was rebuilt around an achromatic-ledger layout:
viewer bar, entrypoint selector, table of contents, summary, survival panel, rotation
plan, cross-scenario matrix, graph, findings, assumptions, data files and diagnostics
in one printable document. - Geist and Geist Mono are embedded as WOFF2 data URIs under the SIL Open Font License
1.1. The report remains fully offline and renders the same typography on every
platform. - The authority graph gained zoom-out, zoom-in and fit controls, off-path dimming,
label backgrounds with ellipsis and a brighter dark-theme edge token. - A boot gate hides the document until the first hydrated render, removing cumulative
layout shift. A no-script override keeps the document visible and shows a notice
that the report needs JavaScript. - Long identifiers wrap at word boundaries, wide tables are keyboard-scrollable with
visible hints, and matrix column labels are shortened while full names remain in the
accessible table headers. scripts/check_artifacts.pynow requires the embedded fonts and their OFL licence
file in the wheel.- The Terraform JSON parser no longer uses an
assert; malformed input returns a
diagnostic on every path. Bandit's B105 skip is recorded with its justification:
the flagged strings are credential-kind names, not credential values.
Verification
Recorded on Windows 11 (10.0.26200, x64), CPython 3.12.14, unless a GitHub Actions run
is cited. The complete release-gate record is in docs/production-readiness.md.
- Full suite: 418 passed, 1 skipped (a real Windows symlink needs an unheld privilege;
Linux CI executes it). Remote CI run 40509205221 passed the package job and all six
test jobs on Ubuntu and Windows for Python 3.12–3.14:417 passed, 2 deselected
plus2 passedfor the slow suite, 0 skipped, coverage 86.73–86.74%. - Adversarial suite: 37 passed, 1 skipped. Golden: 2 passed. Performance: 2 passed
(graph build plus scenario 1.57 s of a 5 s budget; rank 6.29 s of a 60 s budget). - Coverage: 86.67% overall line coverage; every enforced module at or above 80%;
report/html.pyandreport/ui.pyat 100%. ruff check,ruff format --check,mypy --strict srcand
scripts/gen_schema.py --check: clean.- Three consecutive hatchling builds produced identical wheel and sdist SHA-256
digests. The wheel (92 files) and sdist (242 files) passed
scripts/check_artifacts.py. - Clean installs: the wheel with
requirements.lockand the wheel with thehclextra
both ran all six commands; the minimal wheel reportedD-TF-HCL-UNAVAILABLEfor HCL
input while JSON input still worked; the sdist installed and scanned fx01. Repeated
scans produced byte-identical JSON, SARIF and HTML. - Browser QA in headless Chrome: zero axe-core WCAG 2.2 AA violations on desktop dark,
desktop light and mobile dark; zero console errors or warnings; zero external
requests; no horizontal overflow at 1440, 768 or 390 px; fonts loaded and Geist was
actually used; the boot gate clears after hydration. - Lighthouse 13.5.0, desktop profile: performance 98, accessibility 100,
best-practices 100, FCP 0.9 s, LCP 0.9 s, TBT 10 ms, CLS 0. The same report under
simulated mobile throttling scores 65 because the 841 KiB single file is downloaded
at 1.6 Mbps; that profile does not represent opening an offline file from disk. - html-validate 11.17.0: clean. Design inspector: 0 findings at 1440, 768 and 390 px.
- zizmor 1.31.1: 0 findings. actionlint 1.7.12: 0 diagnostics. Semgrep 1.182.1
(p/ci,p/python): 0 findings. Bandit 1.9.4: 0 findings. pip-audit 2.10.1 and
Grype 0.121.0: no known vulnerabilities. pip-licenses: only permissive and
documented weak-copyleft licences, including the fonts' OFL-1.1. - Secret scans: gitleaks 8.30.1 found no leaks across 9 commits and 2.32 MB of
history; TruffleHog 3.99.7 reported 0 verified and 0 unverified secrets. Six
gitleaks tree hits were all triaged to AWS documentation examples inside the
gitignored research cache and are absent from tracked files and artifacts.
Security and assurance limits
The report keeps the same security properties as 0.1.0: restrictive CSP hashing all
five inline scripts, Unicode-escaped embedded JSON, text-only DOM construction, no
network requests and no credential values. Hostile strings are covered by tests and
render inert.
No independent human security audit has been performed, and none is claimed. The
internal audit and the external-audit package define what an independent reviewer
would receive. Cross-machine byte reproducibility and non-Windows/Linux platform
coverage are not claimed.
Upgrade
Replace the 0.1.0 wheel with nhi_blast-0.1.1-py3-none-any.whl; there is no data or
schema migration. Verify the download against SHA256SUMS attached to the release.
NHI-Blast 0.1.0
NHI-Blast 0.1.0
Initial public release, distributed as GitHub Release assets only. NHI-Blast is a
local-first, offline Python 3.12+ command-line tool that reads GitHub Actions
workflows, local composite actions, Terraform JSON/state and IAM policy documents and
answers three incident-response questions from static evidence:
- From a hypothetical compromised entrypoint, which AWS roles, credential stores and
credentials are potentially reachable? - If the entrypoint is revoked, which credentials end with the job (R), which expire
on their own (E), and which survive (Z)? - In what order should surviving credentials be rotated so a surviving credential
cannot re-read a freshly rotated one?
Findings describe potential authority and conditional survival. They do not prove that
a compromise occurred, that a credential was obtained, or that a role can be taken
over.
Install
Download the wheel or sdist and SHA256SUMS from this release and verify the checksum.
The sdist contains requirements.lock; use it to pin runtime dependencies:
sha256sum -c SHA256SUMS
python -m venv .venv
.venv/bin/python -m pip install -c requirements.lock nhi_blast-0.1.0-py3-none-any.whlIf you install from the wheel alone, take requirements.lock from the sdist or the
repository and pass it with -c; the wheel does not embed the lock file.
On Windows use .\.venv\Scripts\python.exe and Get-FileHash. The optional extras are
rich for the enhanced console renderer and hcl for best-effort Terraform HCL
parsing; JSON plan/state input does not require the HCL extra.
What is included
scan,rank,explain,export,schemaandversioncommands with documented
exit codes 0–4.- Offline discovery and parsing of workflows, composite actions, Terraform plan/state
JSON, Terraform JSON configuration and IAM policy documents. - Evidence classes
DECLARED > DERIVED > ASSUMED > UNKNOWN, weakest-class edge
confidence, and a separate potential set for UNKNOWN authority. - Deterministic JSON (
nhi-blast-report/1.0), SARIF 2.1.0 and a single-file offline
HTML report with an exposure matrix, shortest-path trace, scenario-aware summary,
filters and accessibility support. - Rotation planning with containment steps, generation ordering, strongly connected
component batches and verification steps. - Exact dependency locks, an allowlisted artifact checker, three-build digest
reproducibility and a security-marked adversarial test suite.
The full feature list and fixes are in CHANGELOG.md. Rule documentation is under
docs/rules/; assumptions and known limitations are in docs/assumptions.md and
docs/known-limitations.md.
Verification performed for this release
- GitHub Actions CI passed on Ubuntu and Windows for Python 3.12, 3.13 and 3.14, plus a
package job. The run is linked fromdocs/production-readiness.md. - The advisory workflow audits both lock files with pip-audit, runs Bandit, OSV-Scanner
and gitleaks, and runs weekly on a schedule. - The internal audit in
docs/security-audit.mdrecords zizmor, actionlint, Semgrep,
Bandit, pip-audit, OSV-Scanner, Grype, pip-licenses, gitleaks and TruffleHog results,
dependency-age checks, and the fixes made during the audit. - Runtime and development locks were audited with no known advisories on the audit
date; licenses are permissive or weak copyleft only. - The wheel and sdist are byte-identical across three consecutive builds from the same
toolchain, and their SHA-256 digests are published inSHA256SUMS.
Limits of this release
- No independent audit. The project has not been reviewed by an external human
auditor and no independent audit is claimed.docs/external-audit-package.mdlists
what such a review would receive. - Static analysis only. The tool does not read live cloud or GitHub state, and a
finding is not proof of exploitability. - GitHub secret scanning and CodeQL are unavailable on this private repository without
GitHub Advanced Security; compensating scans are recorded in the internal audit. - One symlink security test is skipped on the maintainer's Windows environment because
creating symlinks requires a privilege; the Linux CI job executes it. Windows
directory junctions are always covered. - Cross-machine and cross-toolchain byte reproducibility is not claimed.
License
MIT. See LICENSE and NOTICE. Vendored report assets retain their upstream
licenses, recorded in NOTICE and src/nhi_blast/report/assets/MANIFEST.json.