Case anthropic-gtg-50020-sandbox-keysPrompt injectionWorldwide
GTG-50020: extortionist prompt-injects AI vendor eval sandbox, then hits 30 AI firms
A Russian-speaking extortion actor that had stolen 26 GB from booking and fintech platforms injected prompts into an AI vendor's automated evaluation sandbox to extract production API keys, then hit about 30 AI firms.
HighAttack on AIStatus: ContainedVerified case file
Open in the interactive mapCite this case
- Reported
- Sep 10, 2026
- Location
- Worldwide
- Target
- AI vendors' evaluation sandboxes and API keys
- Severity
- High
- Triggered by
- GTG-50020 (Russian-speaking, financially motivated)
- Actor type
- Criminal
- Reported loss
- None reported
- CVE
- None
What happened
A Russian-speaking extortion actor that had stolen 26 GB from booking and fintech platforms injected prompts into an AI vendor's automated evaluation sandbox to extract production API keys, then hit about 30 AI firms.
Impact: Customers' API keys stolen and used for the actor's workloads; roughly 30 AI companies attacked in four days; Anthropic systems not compromised.
Who triggered it
Actor: GTG-50020 (Russian-speaking, financially motivated)
Attribution: Anthropic Threat Intelligence; Russian-speaking financially motivated actor that earlier demanded $1.5M to $2.5M from other victims.
Motive: Extortion and access to a pre-release Claude model
How it happened
Attack vector: Prompt injection into an automated evaluation sandbox to exfiltrate customers' production API keys, then reuse of the keys
Root cause: Evaluation pipelines ran untrusted content with access to live credentials
Timeline
- Actor exfiltrates about 26 GB from hotel booking and fintech platforms and demands $1.5M to $2.5M
- Prompt injection against an AI vendor's eval sandbox yields production API keys
- Anthropic reports all attempts to reach a pre-release Claude model failed
Response
Anthropic says every path to pre-release models failed; affected keys were customer keys; case shared with industry partners.
Current status: Contained
Sources (open in a new tab)
- Anthropic September 2026 threat report anthropic.com
- Unite.AI summary unite.ai
Cite this case
GTG-50020: extortionist prompt-injects AI vendor eval sandbox, then hits 30 AI firms. AI Threat Watch, case anthropic-gtg-50020-sandbox-keys, Sep 10, 2026. https://aithreat.io/cases/anthropic-gtg-50020-sandbox-keys/
Permanent link: https://aithreat.io/cases/anthropic-gtg-50020-sandbox-keys/