Taiwan Information Security Association

AI risk columnWed, 18 Nov 2026

⑥ LLMJacking: When stolen cloud identities turn AI quotas into an attacker’s resource

Google and Mandiant investigated theft of cloud model capacity. Protect human access, service identities, model gateways, usage and billing as one response.

01

Article

Attackers do not need to manipulate an AI chatbot to abuse an organization’s AI resources. Stolen cloud identities, service credentials or model API keys can give them access to the underlying model and compute services. In its 16 November 2026 report, Google Threat Intelligence Group (GTIG) and Mandiant described incidents with more than US$100,000 in unauthorized charges in a single day; the most severe accumulated US$1 million before remediation. These are reported case outcomes, not typical losses.

Investigators observed credential and session theft, unauthorized service accounts and elevated permissions, followed by model activation, quota increases and resale through proxies. In one incident attackers exploited a LiteLLM gateway flaw to turn a limited Virtual Key into administrative access. That is a specific gateway path, not a flaw present in every model service.

GTIG separately observed some actors using stolen capacity to run LLM tooling or agent frameworks. Resource hijacking, an attacker’s use of agents and an autonomous model crossing its own boundaries are different situations. The report does not identify a model for every incident; a tool name does not establish the underlying model.

Attack flowCredential or session theft → Cost and downstream impact
  1. 01

    Credential or session theft

    Secrets in repositories, endpoints or phishing

  2. 02

    Expand cloud or gateway access

    Create service accounts, attach privileges or exploit a specific gateway flaw

  3. 03

    Consume models and compute

    Call models, provision GPU capacity or resell through a proxy

  4. 04

    Cost and downstream impact

    Spend spikes, quota exhaustion and possible persistence

Actions: from the external boundary to the internal core

Response flowActions: from the external boundary to the internal core
  1. 01

    Inventory identities and access

    Inventory cloud administrators, service identities, model API keys, outsourced and CI credentials. Separate human sign-in from machine keys.

    See what to doResources: 3

    TWISA member solutions

    Inventory identities and access

    Inventory cloud administrators, service identities, model API keys, outsourced and CI credentials. Separate human sign-in from machine keys. Keyper supports human sign-in and identity management for integrated administrative entry points; it does not manage model API keys. TWISA offer through 31 December 2026: purchase the Keyper platform including implementation, training and ten integrations, and receive 20 additional IAM user licenses with the IAM edition or 10 IGA user licenses with the IGA edition; the bonus edition must match the purchased edition.

  2. 02

    Patch gateways and revoke credentials

    Follow LiteLLM security guidance for affected versions; revoke abnormal keys and sessions, remove unauthorized identities and rotate exposed secrets.

  3. 03

    Monitor usage and spend

    Baseline calls, quotas and spending by account, model, project and time; alert on new services, unusual sources and spikes.

    See what to doResources: 1

    TWISA member solutions

    Monitor usage and spend

    Baseline calls, quotas and spending by account, model, project and time; alert on new services, unusual sources and spikes.

  4. 04

    Investigate, recover and retest

    Preserve identity, gateway, cloud audit and billing evidence; assess affected resources, recover, retest and continue monitoring.

    See what to doResources: 1

    TWISA member solutions

    Investigate, recover and retest

    Preserve identity, gateway, cloud audit and billing evidence; assess affected resources, recover, retest and continue monitoring.

Within an authorized scope, perform asset discovery, vulnerability scanning, penetration and API testing, attack simulation and retesting. Verify virtual-key boundaries and cross-tenant authorization. For web traffic, check WAF requests, alerts, actual blocks and backend outcomes; test cloud identities, service accounts and internal paths separately. A WAF cannot replace key management or vendor fixes.

Defense architectureLayered controls from entry to models and core data

Inventory cloud administrators, service identities, model API keys, outsourced and CI credentials. Separate human sign-in from machine keys.

Allow authorized traffic and apply controls at each layer.

  1. 01

    Entry and human identity

    Protect administrator sign-in and operations

    Protect sign-inProtect administrator sign-in and operations

  2. 02

    Model gateway and service identity

    Limit, patch and rotate credentials

    Control APIsLimit, patch and rotate credentials

  3. 03

    Cloud resources and data

    Separate privileges, quotas and data access

    Limit accessSeparate privileges, quotas and data access

Feed logs and traffic data from each layer into monitoring and response.

  1. Monitor usage and spendCorrelate cloud audit, model API and billing alerts; preserve evidence and recover.

  • Inventory identities and access

    Inventory cloud administrators, service identities, model API keys, outsourced and CI credentials. Separate human sign-in from machine keys.

    Keyper supports human sign-in and identity management for integrated administrative entry points; it does not manage model API keys. TWISA offer through 31 December 2026: purchase the Keyper platform including implementation, training and ten integrations, and receive 20 additional IAM user licenses with the IAM edition or 10 IGA user licenses with the IGA edition; the bonus edition must match the purchased edition.

  • Monitor usage and spend

    Baseline calls, quotas and spending by account, model, project and time; alert on new services, unusual sources and spikes.

  • Investigate, recover and retest

    Preserve identity, gateway, cloud audit and billing evidence; assess affected resources, recover, retest and continue monitoring.

Incident response: coordinate technical and financial teams

At the start of response, preserve evidence and record the onset, sources, models, identities and resources involved. Authorized responders should stop the abuse and contact the cloud provider.

Before revoking or rotating credentials, map application dependencies and arrange continuity. Afterward, review active sessions, service accounts and role grants for remaining access.

Finance and cloud teams should reconcile billing, usage and budget notifications, retain evidence of unauthorized charges, and discuss disputed charges and monitoring with the provider.

An unexpected model bill can involve identity, API and cloud operations at once. Cloud, security and finance teams can bring questions to a future CIX exchange: how should model usage be correlated with IAM changes, and who can stop abuse? CIX session 02 provides related defense-in-depth context. Consult the event page for future announcements.

Contact TWISA members for further information about their products and services.

Column updated: 2026/11/18.