Ledger / 10 Sep 2026
GTG-50020 injected prompts into an AI vendor sandbox to steal API keys
Confirmed
Anthropic reports from its own investigation.
The actor used stolen keys and AI pentest loops against victims and about thirty AI firms.
“By injecting malicious instructions into an AI vendor's automated evaluation sandbox, the actor caused the sandbox to hand over the credentials it held”Anthropic
Sources
- Detecting and countering misuse of AI: September 2026Anthropic · 10 Sep 2026 · primary source
- Disclosed
- 10 Sep 2026
- Activity
- 2026
- Actor
- GTG-50020 (criminal), as attributed by Anthropic
- Nexus
- Russia (Russian-speaking)
- Targets
- Hotel booking and fintech platforms; about thirty AI companies
- What the AI did
- Agentic intrusion, Ops automation, Exploit dev, Agent hijack, LLMjacking, Supply chain
- Model
- Claude
- Model family
- Anthropic
- Model basis
- stated by vendor
- Autonomy
- semi-autonomous
- Outcome
- About 26 GB stolen from one victim; AI keys stolen; pre-release model access never obtained.