The attacker did not need to compromise the underlying model weights. Stolen API credentials exposed a billing and capability surface, and the operator routed requests through a service whose advertised token prices made compromised access commercially attractive. Google ties a specific rapid volume of requests to a victim billing exposure. The monetary…
Inspect the ClaimsCase · DiggingBeagle record
Stolen AI API access supported more than 500,000 model calls in eight hours
Google described a token-reseller operation that passed substantial model charges to an account whose credentials were stolen.
Public disclosure: Nov 5, 2026
Scope
Observed stolen-key abuse against hosted language-model APIs and token-reseller infrastructure reported in November 2026.
Assessment profile
Assessment method- Setting
- production
- Exploitation
- confirmed in the wild
- Evidence
- Grade B
- Remediation
- mitigation available
- AI role
- AGAINST AI
Basis and provenance
Google Threat Intelligence reported more than 500,000 model API calls from stolen third-party keys in under eight hours; loss and criminal revenue cannot be equated. An impact band remains unassigned because the available evidence does not establish a comparable severity score or verified numerical loss for this assessment.
Assessed Nov 8, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-GOOGLE-LLMJACKING-TOKEN-RESELLER-01 · CLM-ENR-JETNIGHT-RESELLER · Google Threat Intelligence: LLMJacking monetizes compromised AI resources
At a glance
Hosted model access is both a compute capability and a billable credential. A low advertised resale price can be an indicator of stolen access, but not proof for a specific provider without usage evidence.
Read the implicationsThe monetary figure is the reported usage charge associated with stolen keys, not the actor’s net revenue or proof of completed customer payments. The intelligence report does not provide a universal incidence rate across cloud providers.
Limits and uncertaintyFull account
The attacker did not need to compromise the underlying model weights. Stolen API credentials exposed a billing and capability surface, and the operator routed requests through a service whose advertised token prices made compromised access commercially attractive. Google ties a specific rapid volume of requests to a victim billing exposure.
The monetary figure is the reported usage charge associated with stolen keys, not the actor’s net revenue or proof of completed customer payments. The intelligence report does not provide a universal incidence rate across cloud providers.
Google describes a token-routing business model, not a model jailbreak. Stolen third-party API keys let the Jet Night service pass more than 500,000 model calls in less than eight hours to a victim account, with nearly US$1 million in reported token charges. Google also reported inexpensive customer subscription offers, but the usage charge cannot be read as the attacker's collected revenue. Detecting the abuse requires billing, model-route and token-identity telemetry rather than monitoring the words in model responses.
Mechanism and trust boundary
- 01
Obtain cloud or third-party model API credentials.
- 02
Route outside customer requests through the compromised account.
- 03
Accumulate model usage charges at high volume.
Claims & evidence
CLM-ENR-JETNIGHT-RESELLERGoogle identified the service Jet Night as an example of stolen-key token routing and described an advertised unlimited-token offer near US$60 per month, without establishing the operator’s collected revenue.supported
Basis: reported finding
- supportsGoogle Threat Intelligence: LLMJacking monetizes compromised AI resourcesprimary disclosure
LLMJacking monetization and Jet Night token routing service paragraph
CLM-DB233-GOOGLE-LLMJACKING-LIMITGoogle Threat Intelligence reported inference theft measured in requests and estimated provider charges; these figures do not establish collected reseller revenue or recovered victim payments.supported
Basis: reported finding
- supportsGoogle Threat Intelligence: LLMJacking monetizes compromised AI resourcesprimary disclosure
Incident cost calculations, cloud misuse and resale discussion
CLM-GOOGLE-LLMJACKING-TOKEN-RESELLER-01Google Threat Intelligence observed a threat actor making more than 500,000 model API calls within eight hours using stolen third-party keys, with victim costs approaching one million US dollars.supported
Basis: reported finding
- supportsGoogle Threat Intelligence: LLMJacking monetizes compromised AI resourcesprimary disclosure
Jet Night token routing service and eight-hour API-call count
Implications
Hosted model access is both a compute capability and a billable credential. A low advertised resale price can be an indicator of stolen access, but not proof for a specific provider without usage evidence.
Controls and mitigations
- Scope, rotate and budget-limit model keys.
- Alert on unexpected token volume and model routes.
Unknowns and contradictions
- The monetary figure is the reported usage charge associated with stolen keys, not the actor’s net revenue or proof of completed customer payments. The intelligence report does not provide a universal incidence rate across cloud providers.
Sources and citation
Material revision history
- Nov 10, 2026 · Published version · first publication · revision 234