▣ Skill Poison — Knowledge ExplorerONLINE / EDITION 10
AGENT SKILLS / SECURITY
Independent. Open. Curious.
← CASE ARCHIVE
C01 / Published packageJanuary–February 2026

ClawHavoc: prerequisite lures at marketplace scale

Malicious skill listings used ordinary setup language to move the operator toward external malware.

EVIDENCE CLASSPublished package

Reported observations below are attributed to the linked sources. Review questions and controls are editorial interpretation.

What the source reports

Antiy describes malicious ClawHub packages disguised as useful extensions, with ClickFix-style setup lures. Its report credits Koi with naming ClawHavoc. OpenSourceMalware independently documents early uploads using a supposed authentication prerequisite. These accounts can cover overlapping artifacts.

Primary records: S25 ↗, S29 ↗, S30 ↗

The trust boundary

A package author’s setup instructions were treated as a reason to trust an external installation step.

What this does not establish

Reports document packages and delivery behavior. They do not establish how many unique people installed them or suffered verified losses. Different dates and collection methods explain why published totals differ.

We reviewed the linked publication text or cataloged evidence level. We did not execute the suspect package, independently reproduce the investigation, or verify current registry availability.

Questions for a review

These are practical questions prompted by the case, not controls independently validated by this website.

Prerequisite origin
Who owns the alleged dependency, and which evidence independent of the skill establishes its identity?
Agent-to-human handoff
Does the workflow ask a person to perform an action the constrained agent cannot perform?
Counting overlap
Do reports describe the same package revisions, publisher identifiers, or campaign window?

Editorial lesson

Review prerequisites as separate dependencies. The approval to use a skill is not approval to run every installer it recommends.

For your own environment, turn the question into a bounded test with an exact package identity, synthetic input, expected permitted behavior, prohibited effect, and observable stopping point. Keep a blocked attempt distinct from a completed effect. Preserve the legitimate task result alongside the security observation.

Sources & reading path

S25
ClawHavoc: Analysis of Large-Scale Poisoning Campaign Targeting the OpenClaw Skill Market for AI AgentsAntiy CERT · Publisher report checked 15 October 2026
↗
S29
Malicious ClawHub Skills Target OpenClaw UsersOpenSourceMalware research team · Report updated 3 February 2026
↗
S30
From Automation to Infection: How OpenClaw AI Agent Skills Are Being WeaponizedBernardo Quintero · VirusTotal · Publisher report checked 15 October 2026
↗

Understand the mechanism

Record your own review in the workbench →

Search guides, research, and glossary termsESC TO CLOSE

Start typing to explore the knowledge base.