deerwork-ai
Featured
deer-workflow
An open-source Dynamic Workflow runtime that combines deterministic TypeScript orchestration with replaceable Agent runtimes.
AI repository intelligence
FR-AI / ANALYSISWhy it is worth attention
It addresses the hard post-advisory forensics question the advisory does not answer, with an ordered investigation process, a detector that can classify crafted files via ranged reads, and provenance from a real 37signals investigation.
Who it is for
- Rails teams using Active Storage with the vips variant processor
- Security engineers and incident responders needing to assess possible exploitation
- Developers who want to understand the attack chain and investigation method
- Operators using AI coding agents such as Claude Code to run structured skills
Use cases
- Run kr2s-was-i-vulnerable to establish the exposure window after patching
- Run kr2s-was-i-exploited to produce a written forensic analysis when exploitation is suspected
- Scan the Active Storage blob store for crafted MAT files with bin/kr2s_scan_active_storage_blobs.rb
- Read reference/the-attack.md and reference/the-investigation.md to understand the vulnerability and why artifacts remain
Strengths
- Attack chain and investigation are documented call by call with the load-bearing junctions marked
- Detector needs only the first 128 bytes via a ranged read, so candidates can be classified without downloading whole objects
- The production scanner is committed as a single self-contained file requiring no build step
- Tooling came from a real 37signals investigation and correctly identified the security researchers' test files
Considerations
- Only covers Active Storage with the :vips variant processor and tracked variant records; other libvips upload paths need separate sweeps
- A clean scan result is strong evidence, not proof, and the analysis template has a mandatory limits section
- Skills depend on repository-relative shared references, so copying a skill directory alone breaks them; investigation is time-sensitive because logs expire and unattached blobs are purged
README quick start
rails-forensics-CVE-2026-66066
Tooling and method for answering two questions about a Rails application and CVE-2026-66066, the Active Storage and libvips arbitrary file read known as KindaRails2Shell:
- Was I vulnerable, and for how long?
- Was I exploited, and if so, what left the building?
The advisory explains the vulnerability and names the patched versions. It does
not tell you how to find out whether anyone used it against you. That question
decides whether secret_key_base and everything it signs have to be rotated,
and it is harder than it looks. The evidence is spread across your object store
and three Active Storage tables, application logs expire long before the
exposure window does, and a scheduled purge of unattached blobs destroys
precisely the records the answer depends on.
If you are going to ask, ask soon. Every one of those is a clock that is already running.
This is a documentation and skill repository rather than a library. Nothing here is installed into your application. An agent reads it and works against your application checked out somewhere else, and the one script that runs in production is a single self-contained file you copy into a container.
The vulnerability in one paragraph
Rails decides a blob may be image-processed by reading a content_type column
that a direct upload lets the client set without the bytes ever being examined.
libvips then decides what the file actually is by sniffing magic bytes. A file
whose first ten bytes claim MATLAB 5.0 is routed to libvips' MATLAB loader,
which hands it to libmatio, which dispatches on a different byte range and
finds MAT 7.3, which is HDF5. HDF5's External File List lets a dataset's bytes
live in another file named by path and offset, so rendering the "image" reads an
attacker-chosen file off the server and returns its contents as pixels. The same
confusion, twice, at two layers that cannot see each other's fields.
reference/the-attack.md traces the whole call chain, from the inbound request
to the point where libmatio opens the target file, with the four load-bearing
junctions marked. reference/the-investigation.md then explains why the attack
is investigable at all: it leaves three artifacts of increasing strength, and
the strongest of them is the rendered variant, which holds the stolen bytes as
pixel values in your own object store.
Start
Related repositories
Similar projects matched by category, topics, and programming language.
gavamedia
Featured
deltafin
Deltafin is a research project that runs the 2.8-trillion-parameter Mixture-of-Experts model Kimi K3 on a single Apple Silicon Mac (e.g., M1 Max with 64 GB) at about 16 seconds per token, using exact, reproducible inference with local or streaming expert loading.
7-e1even
Featured
learn-agent
A collection of notes on building coding agents, derived from the production development of the desktop agent Reina, with each mechanism simplified into a zero-dependency, single-file Node.js demo.