The OAuth Token Exploit
AI accounts are increasingly becoming prime targets for unauthorized inference harvesting. When compute capacity remains scarce and expensive, siphoning another company's enterprise quota represents a lucrative underground business. On August 4, Grant De Swardt, an independent AI consultant based in East Sussex, U.K., noticed unexplained token consumption surging across his Claude Max 20x account. Even after he severed his connected integrations and halted all active workloads, the meter kept running. Anthropic later confirmed that an attacker had weaponized a compromised Claude session key to mint unauthorized Claude Code OAuth tokens in the background.
Support Delays and Tracking Gaps
To contain the breach, Anthropic suspended De Swardt's paid account, invalidated his active sessions and server-side Claude Code tokens, and issued a partial refund of £44.49 for the remainder of his $200-per-month plan.
For engineering leaders deploying automated pipelines, the incident exposes the operational hazard of running LLM integrations without granular telemetry. When platforms provide only aggregated consumption metrics rather than per-client or IP-level audit trails, identifying a runaway leak turns into an agonizing guessing game. Mitigating this shadow inference threat requires immediate operational discipline: isolating development environments, automating strict secret rotation schedules, and instrumenting anomaly-detection alerts for baseline API spikes before providers invoice you for compute you never ran.