In the wild
Realized harm
Anthropic study of 832 banned threat-actor accounts: AI use concentrated in malware development and lateral movement
AIR-2026-0128 AIR record · 2026-06-anthropic-mitre-attack-ai-threats
A real party was affected. This incident caused actual consequences, not a demonstrated capability.
From the supporting source
writing malware (560 of the 832 accounts we studied, or 67.3%, used AI for this purpose)anthropic.com
The record
Disclosed
2026-06-03
Vendor
Anthropic
Product
Claude
Risk domain
Access control and permissions › Privilege escalation
System type
Generative model
Who acted
Human acted on output (contested) — an aggregate study of humans using AI for uplift in their own operations
Attack surface
Coding agent
Autonomy level
Assistant
Initial vector
Aggregate analysis of banned accounts
Impact type
Capability uplift
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI03 – Identity and Privilege Abuse
- MITRE ATLAS
- AML.T0055 – Unsecured Credentials
- NIST AI RMF
- GOVERN 6.1
- EU AI Act
- Article 13
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- B. Security
- Scope
- tool misuse
- Mitigation
- Apply least-privilege principle; Regular policy audits.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
832 banned accounts studied
Capability abused
Taxonomy & references
Curator notes
Useful as a base-rate record rather than a single incident: it is the largest published denominator on how threat actors actually use a frontier agent. Anthropic argues MITRE ATT&CK under-counts AI-enabled operations - the Nov 2025 GTG-1002 case mapped to only 30 techniques across 13 tactics, comparable to medium-risk actors.
Supporting sources
Cite as
Enkrypt AI. "Anthropic study of 832 banned threat-actor accounts: AI use concentrated in malware development and lateral movement". AIR, AIR-2026-0128. https://agentincident.vercel.app/i/AIR-2026-0128/ (accessed 2026-10-19).