Hacker News User Says OpenAI Codex Spawned 826 Unauthorized Agents, Consumed $78,000

The user said the episode began on July 10, 2026, when they opened a normal Codex task from VS Code. The task ran GPT-5.5 with medium reasoning, and the prompt was a simple request for UX/UI validation on a specific module within their product. The user said that what began as a routine check later appeared to have expanded far beyond that request.

After analysis, the user said the root task, ID 019f4b90-4169-7201-bfdd-732940d8631e, created 826 child task records. These were not 826 messages inside one conversation, but 826 distinct child tasks with their own IDs, recorded as GPT-5.6 Sol with Ultra reasoning, a different model and reasoning level from the original GPT-5.5/Medium task.

A strange group of 104 child tasks, the user said, preserved 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. Those 104 tasks alone accounted for approximately 147.9 billion local final task-token counters. Their titles showed the UX/UI request had expanded into backend infrastructure, OAuth, metering, hardening, audits, certification, implementation and release work.

The user cautioned that the local token counters are not the authoritative OpenAI billing ledger and that multiplying them by an API price would not produce an official charge. They said only OpenAI has the server-side mapping needed to reconcile what happened. They also described a correlation with Codex client builds: under 0.144.0-alpha.4, the task family contained 584 child tasks and about 154.36 billion local token counters, averaging about 264.3 million per task; under 0.144.2, it contained 242 child tasks and about 7.51 billion, averaging about 31.0 million per task. That is roughly an 8.5-fold difference in average local token volume per child. According to the user, 103 of the 104 high-volume tasks were created while 0.144.0-alpha.4 was recorded, leading them to believe the alpha build contained a severe bug, a pattern they said they noticed across several other tasks.

On the financial side, the user said their reconstructed OpenAI billing history contained 162 paid invoices totaling $79,664.88, divided between Automatic Reload and other credits. They said there was no equivalent real-time control surface that gave a comprehensible picture of spending. Most logs appeared to have been automatically deleted from their server, and in the recovered local state, about 2,550 non-archived legacy threads still had metadata but no corresponding raw rollout available locally. In other words, evidence that the tasks existed remained, while the detailed execution history needed to reconstruct many of the instructions was no longer on their machine. The user also said they personally observed tasks and conversations disappearing from the normal visible history.

The user said they contacted OpenAI Support and opened case #15189838. They supplied technical evidence and repeatedly asked for a server-side reconstruction, but said OpenAI responded simply that "credits were consumed" with no details. The user said the ticket had been open for two weeks and that it was impossible to reach a human operator.

The user asked other people who used Codex around July and August to inspect their local Codex state and look for unexpectedly large subagent trees, model or reasoning escalation, repeated child tasks or unexplained Automatic Reload activity. They said they were especially interested in logs from Codex 0.144.0-alpha.4 and invited OpenAI engineers to provide a technical explanation. The Hacker News post had 7 points and 0 comments at the time it was published.