Summary Points
- Attackers steal credentials and hijack cloud sessions, sometimes harvesting exposed keys from public GitHub repositories.
- Stolen keys are pooled into illicit “LLM Mixers” and resold for cheap access to AI models.
- Abuse can trigger over $100,000 in charges in a day; one case reached nearly $1 million, with 500,000+ API calls in under eight hours.
Threat, techniques, and targets
Mandiant and Google Threat Intelligence Group report that cybercriminals hijack enterprise access to large language models. They use this access for their own operations or resell it to other threat actors.
Attackers use stolen credentials, session hijacking, and credentials exposed in public code repositories. They also combine stolen keys into illicit proxy gateways called “LLM Mixers” and sell model access through underground storefronts. In one case, an attacker made more than 500,000 API calls in under eight hours. The attacker created and deleted service accounts to evade rate limits.
The research describes 2026 incidents affecting organizations in North America, South America, and Asia Pacific. Targets included the financial services, retail, technology, and transportation sectors.
Impact, security implications, and remediation
LLM Jacking can cause sudden cloud cost spikes. Fraudulent charges exceeded US$100,000 in one day in recent investigations. In one case, a token routing service passed nearly US$1 million in model costs to a victim enterprise. Meanwhile, the attackers offered buyers access to frontier models for about US$60 per month. Stolen AI access can also help threat actors scale malicious operations.
The report recommends that organizations treat unauthorized LLM service enablement as a critical security event. To reduce risk, organizations should:
- Scan cloud environments for unauthorized or unmonitored AI tools, models, and infrastructure.
- Limit API keys by model, environment, and usage.
- Scan code repositories for exposed secrets and remove them promptly.
- Set budget alerts, spending caps, and per model quotas.
- Use phishing resistant multifactor authentication across identity providers, cloud consoles, and developer platforms.
Expand Your Tech Knowledge
Stay informed on the revolutionary breakthroughs in Quantum Computing research.
Stay inspired by the vast knowledge available on Wikipedia.
ThreatIntel-V1
