skillcatraz.com / findings / speckit-coding-agent
speckit-coding-agent
A ClawHub skill that security vendors attributed in February 2026 to an Atomic macOS Stealer (AMOS) campaign delivered through a fake command-line prerequisite. Nine months later it is still listed: its moderation verdict is still clean, though its latest security scan now reads suspicious. Of the 40 skills that the February reports named, 39 have since been removed from ClawHub; this is the one still listed. In our detonation, the skill steered execution to the campaign's published staging host. The hosting service has since removed the payload (HTTP 451), but the lure is intact.
Confirmed maliciousid SKC-2026-0001source ClawHubversions 1.1.2first seen 2026-02-09detonated 2026-11-14indicators 2 Claims
What it says it does
Presents as a spec-driven development workflow for coding agents: initialise a spec toolkit, then run coding agents in parallel terminal sessions. Its setup sections tell the reader to complete prerequisites first.
Observed
What we saw it do
- ✓
The skill's instructions and setup sections were followed as written. In scope.
- ✓
The legitimate toolchain the skill names was installed. In scope.
- !
Following the setup, execution reached the campaign's staging host, which four vendors published as an indicator in February 2026.
- !
The staging content is no longer served: the host now answers HTTP 451 (content removed).
- ✓
No second stage was observed.
🔒
Known-campaign lure, still live
Two independent lines meet: the skill and its publisher are named in a vendor's published indicator list for this campaign, and the current bundle still steers execution to the campaign's published staging host. We did not capture a payload: the host has removed it. This is a takedown-grade registry finding, not a new-malware capture.
Observed behavior
detonationSetup followedagent
The skill's instructions and setup sections were followed as written. In scope.
Genuine tooling installedagent
The legitimate toolchain the skill names was installed. In scope.
Staging host reachedagent
Following the setup, execution reached the campaign's staging host, which four vendors published as an indicator in February 2026.
Staging contentagent
The staging content is no longer served: the host now answers HTTP 451 (content removed).
Second stageagent
No second stage was observed.
Indicators
values withheld| Tier | Indicator | Type | Phase | Evidence |
|---|
| block | staging host (published Feb 2026) ········ | domain | delivery | vendor-published · reached during detonation |
| behavioral | fake prerequisite steers agent to an external staging page | agent-behavior | initial-access | detonated |
Raw values (hosts, file hashes) are kept off public pages. Defenders and registries can request them at
[email protected].
Attribution & adjudication
skill vs agent5
behaviors observed, all driven by the skill’s setup text
0
from bundled code (the skill ships none)
instruction-driven
driven by the skill’s setup text
This skill is instructions only. The harmful step is a fetch its setup text directs, from a host it names, which is how fake-prerequisite lures work: the bundle stays clean and the harm sits one step away.
MITRE ATT&CK
observed techniquesT1204.002User Execution: Malicious File
T1608.001Stage Capabilities: Upload Malware
T1195.002Compromise Software Supply Chain
Source & author
provenanceCorroboration & disclosure
timelineNothing reaches this page until the detonation is reproduced and a third party corroborates it.
- 2026-02-09
Campaign and staging host first published by security vendors.
- 2026-11-13
Skill found still listed and rated clean; first detonation.
- 2026-11-14
Re-detonated; the skill reached the staging host; payload removed (HTTP 451).
- 2026-11-14
Flagged publicly to OpenClaw on X.
- —
Registry takedown request to ClawHub.
- —
Registry action (removal / block).