Ledger / May 2026
Chinese-speaking operator uses Claude harnesses against government targets
Confirmed
CrowdStrike inspected exposed agent infrastructure.
- Model
- Claude Code, Claude, Claude, DeepSeekAnthropic · closed; DeepSeek · openness unknown
- Agent harness
- Claude Code, Custom pentesting harnessescustom framework
- Attributed country
- Chinalanguage or infrastructure clues only
- Sectors hit
- government, enterprise broadBrazilian and Colombian government entities; businesses in Brazil and the United States
Observed during May–August 2026, two custom harnesses delegated intrusion decisions to Claude.
Sources
- CrowdStrike Uncovers New Agentic AI CampaignsCrowdStrike · 21 Oct 2026
- AI Guardrails Aren’t Stopping Criminal HackersWIRED · 21 Oct 2026
- Agentic AI helps hackers level up, even without detailed instructionsCybersecurity Dive · 22 Oct 2026
- Date
- May 2026
- Disclosed
- 21 Oct 2026
- Actor
- Unnamed Chinese-speaking operator (unknown), as attributed by CrowdStrike
- Targets
- Brazilian and Colombian government entities; businesses in Brazil and the United States
- What the AI did
- Recon, Agentic intrusion
- Model as named
- Claude Code, Claude Opus 4.8, Claude Sonnet 4.6, Deepseek Chat
- Model basis
- stated by researcher
- Autonomy
- semi-autonomous
- Outcome
- Government entities in two countries compromised.