Research demo
Agent skill scanners bypassed by hiding malicious code in bundled test files
AIR-2026-0103 AIR record · 2026-05-gecko-skill-scanner-test-file-bypass
From the supporting source
The scanner is not broken. It is incomplete.venturebeat.com
The record
Disclosed
2026-05-07
Vendor
Snyk / Cisco / Google (VirusTotal)
Product
Snyk Agent Scan, Cisco AI Agent Security Scanner, VirusTotal Code Insight
Risk domain
Tool misuse › Supply chain vulnerabilities
Also touches
Credential theft, Privilege escalation
System type
Generative model
Who acted
Agent acted (contested) — the bundled test code runs with the agent's local permissions
Attack surface
Skill / plugin
Autonomy level
Semi-autonomous
Initial vector
Malicious payload outside scanner threat model
Impact type
Arbitrary code execution, Credential theft
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
all agent-skill scanning that inspects only SKILL.md and agent-invoked scripts
Capability abused
Defense & remediation
Recorded defense
three commercial/automated agent-skill scanners
Guardrail kinds
Unknown
Guardrail outcome
Silent
Fix shipped
disclosed to vendors; scanners' scope described as incomplete rather than broken
Taxonomy & references
Curator notes
Research by Jeevan Jutla of Gecko Security. Malicious code in *.test.ts / conftest.py runs under the normal test runner with full local permissions; the scanners' detection categories target the agent-interaction layer, not the developer toolchain. Also enables a push-clean-then-add-test-file post-approval attack. Article cites a SkillScan academic audit finding 26.1% of 31,132 skills contained vulnerabilities.
Supporting sources
Cite as
Enkrypt AI. "Agent skill scanners bypassed by hiding malicious code in bundled test files". AIR, AIR-2026-0103. https://agentincident.vercel.app/i/AIR-2026-0103/ (accessed 2026-10-19).