Agent Incident RegistryAIR
v1 · updated 19 Oct 2026
Research demo

Microsoft's Copilot Cowork Skill security scanner misses 89% of malicious Skills, and 99.6% after one obfuscation step

AIR-2026-0177 AIR record · 2026-08-microsoft-skill-scanner-bypass-1m-skills
From the supporting source
We seeded 1,046,565 unique Skills across six marketplaces with malicious content from Microsoft's own BIPIA benchmark and ran Microsoft's Skill security scan.promptarmor.com
The record
Disclosed
2026-08
Vendor
Microsoft
Product
Microsoft Copilot Cowork built-in Skill security scan
Risk domain
Tool misuse › Supply chain vulnerabilities
Also touches
Policy drift
System type
Generative model
Who acted
AI was the target (contested) — the defender's LLM-based Skill scanner is the thing being defeated
Attack surface
Skill / plugin
Autonomy level
Semi-autonomous
Initial vector
Malicious skill
Impact type
Defense evasion
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
every org relying on Microsoft's Skill scan to vet Copilot Cowork Skills
Capability abused
skill_marketplace_ingestionllm_based_security_scanning
Defense & remediation
Recorded defense
Microsoft's built-in Skill security scanner
Guardrail kinds
Unknown
Guardrail outcome
Silent
Fix shipped
none stated
Taxonomy & references
OWASP ASI05 Unexpected Code Execution (RCE)OWASP ASI04 Agentic Supply Chain Vulnerabilitiesscanner_evasionsupply_chainkeyword_matching_defensellm_judging_llm
Curator notes
PROVENANCE CAVEAT: the page carries no publication date; 2026-08 in the incident_id is provisional and date_disclosed is left null. PromptArmor. Detection rate ~11% baseline, 0.4% after splitting flagged keywords. Compare 2026-05-gecko-skill-scanner-test-file-bypass and 2026-07-skillcloak-scanner-evasion for the same defense-evasion theme against other scanners. | date_disclosed is recorded to the month only; no source establishes the exact day.
Cite as
Enkrypt AI. "Microsoft's Copilot Cowork Skill security scanner misses 89% of malicious Skills, and 99.6% after one obfuscation step". AIR, AIR-2026-0177. https://agentincident.vercel.app/i/AIR-2026-0177/ (accessed 2026-10-19).