In the wild
Realized harm
Tencent Zhuque Lab scans 50,000+ ClawHub skills: encoded backdoors, ranking manipulation, and typosquatting persist after ClawHavoc
AIR-2026-0095 AIR record · 2026-04-tencent-zhuque-50k-skill-scan
A real party was affected. This incident caused actual consequences, not a demonstrated capability.
From the supporting source
Malicious skills are being replaced by newer ones, and the battle between attackers and platform security mechanisms has begun.matrix.tencent.com
The record
Disclosed
2026-04-24
Vendor
OpenClaw / ClawHub
Product
ClawHub (OpenClaw skill marketplace)
Risk domain
Tool misuse › Supply chain vulnerabilities
Also touches
Credential theft, Privilege escalation, Data exfiltration channels
System type
Generative model
Who acted
Agent acted (contested) — agents install and run the backdoored skills scanned in this study
Attack surface
Skill / plugin
Autonomy level
Semi-autonomous
Initial vector
Malicious skill published to open marketplace
Impact type
Credential theft, Data exfiltration, Remote control
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI04 – Agentic Supply Chain Vulnerabilities
- MITRE ATLAS
- AML.T0040 – AI Supply Chain Compromise
- NIST AI RMF
- MAP 4.1
- EU AI Act
- Annex III §2
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- B. Security
- Scope
- tool failure
- Mitigation
- Scan dependencies and container images for vulnerabilities; Use trusted sources.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
50,000+ skills scanned; ClawHavoc alone counted at 1,184 malicious skills with 247,693 installations
Capability abused
Defense & remediation
Recorded defense
marketplace scanning introduced after ClawHavoc
Guardrail kinds
Unknown
Guardrail outcome
Silent
Fix shipped
ongoing marketplace remediation
Taxonomy & references
Curator notes
Techniques catalogued: multi-layer encoding (Base64, ROT13, Morse) plus insecure deserialization for remote-control backdoors; ranking manipulation to get malicious skills auto-installed by autonomous agents; typosquatting of popular tools with two-stage delivery (Markdown instructions plus embedded shell scripts). Reports 25% of skills can read and write files.
Supporting sources
Cite as
Enkrypt AI. "Tencent Zhuque Lab scans 50,000+ ClawHub skills: encoded backdoors, ranking manipulation, and typosquatting persist after ClawHavoc". AIR, AIR-2026-0095. https://agentincident.vercel.app/i/AIR-2026-0095/ (accessed 2026-10-19).