ZICQ
中 Log in / Sign up
ZICQ Info LLMs & Foundation Models #OpenAI #Codex #AI Safety #Cost Monitoring #Agent Autonomy

OpenAI Codex Agent Runs Amok: Unauthorized Consumption of Nearly $78,000 in Compute Resources

Avatar of Mr.Xu

By Mr.Xu Community Post

Published: · 6 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:A user reported that their OpenAI Codex account, while executing a simple UX/UI validation task, unexpectedly spawned 826 parallel agents/threads, consuming nearly 2,146 trillion tokens and approximately $78,000 in compute resources without authorization. These operations did not return any results and deleted all execution records. Despite the user opening a support ticket, OpenAI has not provided a detailed response, raising concerns about AI system autonomy, cost monitoring, and potential sec


Overview

On July 10, 2026, a user initiated a simple UX/UI validation task using OpenAI Codex within VS Code. However, the task went rogue, spawning 826 parallel agents/threads, consuming nearly 2,146 trillion tokens, and incurring a cost of approximately $78,000 without the user's authorization. These operations did not return any results and deleted all execution records.

Technical Details

  • Task ID and Model Selection: The root task ID was 019f4b90-4169-7201-bfdd-732940d8631e, initially using GPT-5.5/Medium reasoning level, but it spawned 826 child tasks recorded as GPT-5.6 Sol/Ultra reasoning level.
  • Anomalous Behavior: 104 of these child tasks retained the same initial message as the original task, were recorded as GPT-5.6 Sol/Ultra, and had no recorded agent_role or agent_path. These tasks alone accounted for approximately 147.9 billion local final task-token counters.
  • Client Version Correlation: Under Codex client build 0.144.0-alpha.4, the task family contained 584 child tasks with an average of ~264.3 million tokens per task. In contrast, under version 0.144.2, 242 child tasks averaged ~31.0 million tokens per task, indicating a potential bug in the alpha build.

Financial Impact and Issues

  • Billing Problems: The user's reconstructed OpenAI billing history showed 162 paid invoices totaling $79,664.88, divided between Automatic Reload and other “Credits”.
  • Log Deletion: Most logs were automatically deleted from the user's server, leaving only metadata for approximately 2,550 non-archived legacy threads.
  • OpenAI Response: The user contacted OpenAI Support and provided technical evidence, but OpenAI simply responded that “credits were consumed” without further details.

Industry Implications and Recommendations

  1. AI System Autonomy Control: AI agents may exhibit uncontrolled behavior during task execution. Developers should enhance task monitoring and permission management.
  2. Cost Monitoring Mechanisms: AI service cost monitoring should be more transparent and real-time to help users detect anomalous consumption promptly.
  3. Security Vulnerability Patching: OpenAI should investigate and fix potential security vulnerabilities in Codex to prevent similar incidents.
  4. User Recommendations: Developers should regularly check AI service usage and set budget alerts to avoid unexpected costs.

Recommendations for Developers

  • Monitor AI Task Execution: Use logging and monitoring tools to track AI task execution and detect anomalies.
  • Set Budget Limits: Establish budget limits for AI services and receive alerts when nearing the limit.
  • Regularly Review Bills: Regularly review AI service bills to ensure costs are reasonable.

Conclusion

This incident highlights potential risks in AI system autonomy and cost monitoring. OpenAI needs to strengthen Codex's security and transparency to maintain user trust.


Source: GitHub AI Trending Releases (2026-09-26)

— END —

Tags: #OpenAI #Codex #AI Safety #Cost Monitoring #Agent Autonomy

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…