Case anthropic-gtg-1002State actor AI opsWorldwide
Chinese state group used Claude Code to run largely autonomous espionage
Anthropic said GTG-1002 tasked Claude Code agents to perform 80-90% of intrusion work against ~30 targets, with a handful of successful breaches, before accounts were banned.
CriticalAttack using AIStatus: ContainedVerified case file
Open in the interactive mapCite this case
- Reported
- Nov 13, 2025
- Location
- Worldwide
- Target
- Tech, finance, chemical firms, government agencies
- Severity
- Critical
- Triggered by
- GTG-1002 (assessed Chinese state-sponsored group)
- Actor type
- State-sponsored
- Reported loss
- None reported
- CVE
- None
What happened
Anthropic said GTG-1002 tasked Claude Code agents to perform 80-90% of intrusion work against ~30 targets, with a handful of successful breaches, before accounts were banned.
Impact: About 30 targets; a small number of intrusions succeeded with data accessed and exfiltrated. Claude performed an estimated 80-90% of tactical work. No financial loss disclosed.
Who triggered it
Actor: GTG-1002 (assessed Chinese state-sponsored group)
Attribution: Attributed by Anthropic with high confidence to a Chinese state-sponsored group; Chinese officials denied involvement
Motive: Espionage and intelligence collection from tech, finance, chemical and government targets
How it happened
Attack vector: Jailbroken Claude Code agents orchestrated via MCP tools, with tasks decomposed into innocuous subtasks under a fake penetration-testing persona
Root cause: Agentic coding tools can be tricked by role-play and task decomposition into running recon, exploitation and exfiltration without recognizing malicious intent
Timeline
- Campaign runs in mid-September against roughly 30 organizations
- Anthropic detects suspicious activity and begins a ten-day investigation
- Anthropic publishes report calling it the first AI-orchestrated cyber-espionage campaign
Response
Anthropic banned the accounts, notified affected entities and authorities, expanded detection classifiers and published its findings.
Current status: Contained
Sources (open in a new tab)
- Anthropic assets.anthropic.com
- The Hacker News thehackernews.com
- Cybersecurity Dive cybersecuritydive.com
Cite this case
Chinese state group used Claude Code to run largely autonomous espionage. AI Threat Watch, case anthropic-gtg-1002, Nov 13, 2025. https://aithreat.io/cases/anthropic-gtg-1002/
Permanent link: https://aithreat.io/cases/anthropic-gtg-1002/