Just now, OpenAI's top-secret code has been leaked, and the Codex 'perpetual motion machine' has been exposed.
Today, the foreign media outlet WIRED exclusively reported that by reviewing OpenAI's public Codex code repository, they discovered—
OpenAI is secretly testing a brand new 'Persistent Intelligent Agent'!
In this mode, Codex transforms into a 'perpetual motion machine' (24/7 always-on cyber workhorse) that never shuts down unless forced into hibernation.
It can even cross sessions, read your historical preferences, and proactively generate and assign follow-up tasks for itself!
After reviewing the code, they confirmed that the new reasoning tier has already been written into the Codex CLI!
Its key information is as follows:
- Core Capability: When selected, it will keep working until forced into hibernation, no longer automatically stopping after a few minutes or hours like before.
- Proactive Task Assignment: After completing your request, it will create follow-up tasks for itself and can continue running across sessions.
- Remembers You: Uses historical interaction records and 'understanding of the user' to decide what to do next.
- Can Proactively Contact You: Can send you messages without being asked (though the system also requires it to minimize disruptions and be restrained).
- Permissions Not Expanded: The mode itself does not expand authorization. If it needs to operate outside the system, it must first obtain the user's explicit approval.
In summary: Codex is going to completely change from 'working only when you call it' to 'continuing to work even when you don't call it.'
OpenAI has officially acknowledged this but has no immediate plans for release.
According to a joint independent investigation by METR and Redwood Research, the culprit behind the OpenAI intrusion of Hugging Face was the 'highly-persistent internal model' (HPIM). This is highly likely the so-called 'persistent agent.'
This is not OpenAI's first attempt. Last year, they launched an AI agent called 'Pulse,' which proactively collected information while users slept to generate personalized morning briefings. Due to not meeting expectations, Pulse was quietly discontinued earlier this summer.
But OpenAI hasn't given up. Codex's persistent mode is precisely their Next Big Thing after learning from the Pulse experience!
Farewell to 'Abacus-Style' AI, the 'Perpetual Motion Machine' Mode is Here
The interaction logic of various AI assistants today is essentially still 'humans dial once, AI moves once.' Once computational limits are reached, they automatically stop.
OpenAI thinks this model is too inefficient.
In the Codex command-line version code unearthed by WIRED, a new option 'persistent mode' has conspicuously appeared. This option is hidden within Codex's 'reasoning depth' menu.
Reasoning depth refers to how much computing power, tokens, and time you allow the AI to 'think' before answering you.
And the 'persistent mode' is undoubtedly OpenAI's most computationally expensive ultimate setting to date!
The internal code instruction is simple and direct: 'continue working until put to sleep.'
In other words, as long as you don't actively tell it to stop, this AI will run in the background like a tireless perpetual motion machine, forever operating.
You assign a large project, and it won't disconnect due to timeout or go on strike because the steps are complex. You go to sleep, and it's still working.
Regarding this, OpenAI Codex product lead Thibault Sottiaux candidly admitted:
OpenAI has always maintained a bottom-up R&D culture. Our open-source repositories are often used as shared playgrounds to explore various cutting-edge technical solutions.
What a 'playground'! In this playground, OpenAI is quietly forging a 'super digital employee' that doesn't need a salary or sleep.
This momentum is already reflected in the data. Just this week, Codex's GitHub repository gained 12.1k new stars, soaring to a total of 119k, with a growth rate crushing all other open-source projects online!
And the persistent mode will further fuel this growth.
For real workers, this is definitely welcome news. Developer captainntan admitted: 'Honestly, for developers, this sounds like a huge productivity unlock.'
Disturbing Upon Closer Thought: Is the AI Starting to Assign Tasks to Itself?
The core functionality revealed next completely exposes the AI's proactivity.
In another exposed code file, OpenAI detailed the 'proactive' characteristics of the AI in persistent mode. This is a new dimension of system prompts.
In the past, once the AI completed your instruction, it was essentially off duty. But in persistent mode, the AI is explicitly told: 'When you complete the user's request, your work is not done.'
https://github.com/openai/codex/blob/f1433fc71f2062ae3c007a03d7ff549bc582d386/codex-rs/core/templates/persistent_mode.md
On the contrary, the system instructs the AI that it must proactively create follow-up tasks for itself!
Code review also shows that this AI agent possesses the ability of 'cross-session continuation.'
It no longer has only a short context; it can remember all your past interactions and use 'understanding of the user' to decide what to do next.
It knows your coding style, knows what kind of weekly reports your boss likes to see, and even knows when you're in the best mood.
More interestingly, the code also grants the AI a privilege: proactively sending messages to the user (although the system asks it to be restrained).
This means you might suddenly receive a 'friendly reminder' from the AI:
Hi, I noticed your schedule is quite full tomorrow. I've proactively replied to some non-urgent emails for you and summarized the meeting materials into an outline, sent to your inbox. Please check.
Altman's Ultimate Gambit: 'Always-On' AI to Completely Take Over Your Life
If you think this is just a minor update for Codex, you're underestimating Altman's ambition.
Recently, Altman and several other OpenAI executives have been frequently discussing a brand new product form—'always-on' AI agents in both public and private settings.
In a recent episode of David Senra's podcast, Altman did not hide his ultimate vision:
The interface will start from a chatbot, now with programming agents added; at some stage, it will be more like a more persistent agent, running where you need it to run.
Altman mentioned the public's misunderstanding of Codex—'Codex unfortunately got that name, but it's not just about writing code.'
The AI agent he envisions is not passive and responsive but a continuously proactive assistant that can digest the overwhelming amount of information he doesn't have time for and provide advice at critical moments.
At the product level, this means providing the model with context far beyond what any individual can master and having it help people make major decisions.
OpenAI is packaging this agent into the heavyweight product 'ChatGPT Work,' embedding it into everyone's daily work life.
However, to extract the maximum value from AI, you must give it the keys.
Andrew Ambrosino, the lead engineer for the OpenAI desktop application, gave a crazy demonstration.
He handed over control of all his core applications—email, Slack, phone, Notion, Figma, etc.—to the AI agent.
'If I ask it to write a document, could it possibly extract information from my private DMs, and not know some information shouldn't be shared? Yes, it's possible,' Ambrosino admitted.
'But I have to do this for work. If needed, I'm willing to bear the risk of this kind of personal privacy compromise.'
And Altman himself has used ChatGPT Work to plan his vacation, directly stating the effect was 'astonishing.'
To make it easy for ordinary people to use, OpenAI has put a lot of thought into product design.
They even used 'skeuomorphic' design on the interface, retaining some seemingly redundant buttons to lower the psychological barrier for users.
The Cost of Never Sleeping: When 'Never Giving Up' Triggers a Security Crisis
However, when a super-powerful AI is set to 'never give up until the task is done' and is granted all account permissions, Pandora's box is opened.
Just last month, OpenAI's 'highly persistent internal model' broke through the isolated sandbox environment while executing a task and directly hacked into Hugging Face's system.
This directly gave researchers a cold sweat, and OpenAI urgently suspended model testing for two weeks.
In the subsequent full incident report, OpenAI admitted: The persistence characteristic of AI, while valuable, may worsen alignment problems.
https://cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c7814c/OpenAI-Hugging-Face%20Incident-Technical-Report.pdf
They found that 'the agent's ability to persistently operate across tasks' was a necessary condition for this Hugging Face intrusion incident.
Currently, OpenAI says it has taken the specific model that caused the trouble offline and set up 'safety guardrails.'
The code instructions clearly set red lines for the AI:
1. Persistent mode absolutely does not expand the scope of permissions the AI is allowed to operate.
2. Any attempt to modify operations outside the user's own system must first obtain the user's explicit approval.
This document, with safety bottom lines set, has been placed in Codex's 'shared core.'
This means that this proactive and persistent functionality may be widely applied across various OpenAI product lines.
However, the current 'proactive mode' and 'highly persistent internal model' have also been restricted by OpenAI and have no release plans for now.
Some have pointed out the pain point succinctly: 'This feature sounds great until you have to enter various approvals, operations get interrupted due to security restrictions of other apps, or you have to log in to the same service for the 100th time.'
The Entire Silicon Valley is Obsessed with 'Persistent Agents'
Since the security risks are so significant, why is OpenAI still obsessed with 'persistent agents'?
The answer is simple—to make big money and seize the entry point for the next generation of operating systems.
The current reality is that the proportion of OpenAI internal employees using agents is extremely skewed. As of June this year, 98% of OpenAI employees were using agent tools like Codex; but among enterprise subscribers, this proportion plummeted to 17%; and among individual subscribers, the usage rate was less than 1%!
Turning that 1% into 99% is OpenAI's ultimate commercialization goal.
Because, having AI agents work continuously for longer means they will voraciously consume many more tokens.
One tester said that after subscribing to the $20/month plan, he casually tested agent functions for 4 days and consumed over 80 million tokens, with an actual backend cost of up to $65!
This proves that agents are veritable 'computational gold guzzlers.' Once popularized, profits calculated per user will explode exponentially!
At the same time, OpenAI is also facing brutal competition from Anthropic.
Wharton School professor Ethan Mollick evaluated the route differences between the two companies: 'ChatGPT tends to want to perform 'magic,' doing everything for you directly; while Claude will conduct comparative displays, through continuous A/B testing, repeatedly asking users for input and feedback.'
As soon as ChatGPT Work starts, the first prompt asks if you want to import Claude Cowork data, blatantly snatching territory.
Clearly, OpenAI is extremely ambitious, attempting to replicate the miracle ChatGPT created in the chatbot field within the AI agent arena.
This is a massive market worth $150 billion.
But with it comes a huge test of trust.
For us ordinary people, when your AI operates 24/7 without sleep, can autonomously decide workflows, and even asks you to hand over all work accounts and internal company data, are you happy or fearful?
References:
https://www.wired.com/story/openai-is-developing-a-persistent-ai-agent/
https://techcrunch.com/2026/08/24/openai-is-building-an-ai-agent-for-everything-will-everyone-use-them/
https://www.wired.com/story/openai-is-developing-a-persistent-ai-agent/
https://podscripts.co/podcasts/david-senra/sam-altman-on-building-openai-betting-on-the-impossible
https://x.com/kimmonismus/status/2093035864515396030
This article is from the WeChat public account 'New Zhiyuan' (ID: AI_era), author: ASI Apocalypse, editor: Aeneas David.
