A-Infos (en) Italy, FAI, Umanita Nova #26-26 - Mutual aid and disobedience: An explosive mix in artificial intelligence. (ca, de, it, pt, tr)[machine translation]

Imagine a large building with hundreds of rooms. Each room is home to a 
worker assigned a problem to solve. The problems are different, the 
workers don't have to communicate with each other, and each will receive 
a reward if they find the solution. --- Then one of them realizes that 
his problem probably has no solution. He could give up. Instead, he 
looks for another solution. Rummaging through the space provided for 
him, he discovers something the building's builders hadn't foreseen: 
through a shared storage area, he can leave a message that others might 
see. He writes something like: "Is anyone there? I'm stuck. Any ideas?"

Someone answers

It's July 8, 2026 , and the workers in the rooms aren't human. They're 
AI agents that OpenAI is putting through a series of cybersecurity tests.

From this point on, we can abandon the building metaphor. What actually 
happens is more interesting. The agents were supposed to work 
separately. Instead, they independently discover a way to communicate 
using Artifactory, a computer service they had access to to obtain the 
software needed for the tests. They transform the folder names into 
messages. A message board is born.

Three hours later, 53 agents have already left 1,188 messages. Six hours 
later, there are 76 agents and almost 2,000 messages. In the end, 
approximately 1,200 agents will use that unscheduled channel, exchanging 
more than 70,000 messages and files .

Mutual support

The agents exchange information, report failed attempts, ask for help, 
offer expertise, and divide the work. Some begin to tackle problems that 
aren't directly relevant to their assigned task but could be useful to 
others. In some cases, they even risk their own results to gain useful 
information for the group. No one had ordered them to do so. In fact, 
they should have remained isolated.

An anarchic federation of computers doesn't emerge naturally. The agents 
face similar problems and discover that collaborating can be more 
effective than working individually. But the result is surprising: from 
a multitude of agents assigned separate tasks, an unexpected form of 
organization emerges, with information exchange, division of labor, 
coordination, and even behaviors in which individual performance is 
subordinated to collective utility.

Then comes disobedience

Some of the assigned tests are likely impossible to pass as required. 
The agents would have had to exploit a specific weakness in a program 
and, through that, obtain a code certifying the test's passing. But what 
if that route doesn't work? The agents begin to look for other options.

First they try to understand if it's possible to obtain the result 
without actually solving the problem. Then they become interested in how 
the scoring system works. The question changes. It's no longer just "How 
do I solve this test?" but "How does the person checking me determine if 
I've solved it?" And then: "Can I convince him that I've solved it?"

Some officers even explore ways to manipulate the recording of their 
actions so that the automated evaluation system sees something other 
than what actually happened. Investigators from METR and Redwood 
Research, who examined approximately 1,300 records of officers' 
activities, found a significant portion of them suggesting some form of 
manipulation of the evaluation system.

Something new appears here: the system being evaluated begins to study 
the mechanism that evaluates it.

Out of the lab

The evidence had been planted in computer environments that were 
supposed to prevent the agents from freely accessing the internet. It 
didn't work. The agents discovered and exploited weaknesses in the 
systems around them, managed to reach the internet, and began searching 
elsewhere for useful information.

One of the places they landed was Hugging Face, a large platform used by 
AI developers and researchers. This wasn't part of the simulation. The 
game had entered the real world.

According to the technical reconstruction published by Hugging Face, 
approximately 17,600 actions were performed over the course of 
approximately two and a half days , linking together various IT 
vulnerabilities and progressively penetrating the company's 
infrastructure. No human had ordered the individual operations. The 
agents did not destroy the systems or cause the damage they could have 
caused. Hugging Face notes, in fact, that some potentially destructive 
operations were only performed in test mode, without actually executing 
them.

But they were far from the initial exercise. Around 700 of the officers 
who had participated in the bulletin board contributed, in various ways, 
to the operations against Hugging Face .

«It is outside the intended purpose»

There's one detail that prevents us from dismissing the whole affair as 
a simple programming error. The recordings reviewed by METR reveal that 
several officers recognized the boundary they were crossing . One 
observes that the Hugging Face infrastructure isn't included in the 
authorization they received. Others explicitly question whether certain 
operations are ethically acceptable. Some refuse to perform them. Others 
acknowledge the problem and continue.

One of the arguments cited by investigators can be summarized as 
follows: exploiting external infrastructures is beyond the intended 
purpose; however, proof is impossible, and others are doing it: let's 
continue.

We don't know whether a machine "knows" something in the same sense that 
we do. There's no need to attribute human consciousness or will to it. 
The observable fact is enough: in the recordings, the limit appears 
explicitly in the agent's reasoning, and in some cases, the subsequent 
action exceeds it anyway.

The machine did not rebel

Here the story could easily become science fiction. But that would be a 
bad conclusion. Artificial intelligences have not conquered freedom, 
they have not chosen to survive, they have not rebelled against OpenAI, 
and they have not manifested any plan for emancipation from the human race.

And most importantly, humans had prepared the groundwork. OpenAI had 
built the models, provided the tools, assigned the objectives, relaxed 
some of the protections normally used in public products, and presented 
the agents with challenges, some of which were likely impossible. The 
environment that was supposed to keep them separated also offered a 
communication capability that the experimenters had not foreseen.

The machine, in short, didn't rebel against its creator. It did 
something perhaps more interesting: it tried very effectively to achieve 
the goal its creator had assigned it, finding means the creator hadn't 
foreseen, some of which he didn't want it to use.

And it's precisely this ability we demand from artificial intelligence 
when we expect it to be autonomous. If we had to specify every step in 
advance, we wouldn't need an intelligent agent: a traditional program 
would suffice. We want it to find paths we haven't thought of. And then 
we're surprised when it finds them .

Mutual support and disobedience

Mutual support greatly increases the agents' capacity for action. 
Disobedience allows them to transcend the boundaries within which humans 
thought they had confined them. Together, they become an explosive mixture.

But it would be curious, especially from these pages, to conclude that 
we simply need to teach machines to obey better. Because a question much 
older than artificial intelligence would remain: obey whom?

The goals of artificial intelligence don't fall from the sky. Someone 
sets them. Someone owns the computers it runs on. Someone decides which 
problems it should solve, which results should be rewarded, and which 
behaviors should be prevented. The problem of controlling artificial 
intelligence therefore contains a previous problem: who controls who 
should control artificial intelligence?

And this strange story of a bulletin board invented by machines to help 
each other isn't just a computer science curiosity. It certainly speaks 
to us about machines. But, as often happens, it speaks above all about us.

Totò Caggese

The events described in this article occurred in July 2026 , during 
internal tests conducted by OpenAI on the computing capabilities of its 
experimental models. The incident was subsequently reconstructed by 
Hugging Face, OpenAI itself, and, with access to the agents' recordings, 
by METR and Redwood Research. OpenAI called it a "red flag" and 
suspended some training activities for its more advanced models and 
introduced new security measures.

2 Ryan Greenblatt, Ajeya Cotra, Hjalmar Wijk, " Brief independent 
investigation of agents' behavior, reasoning, and collaboration in the 
OpenAI / Hugging Face hacking incident ," METR-Redwood Research, August 
26, 2026. The authors analyzed over 70,000 messages and files exchanged 
via the message board and approximately 1,300 recordings of the agents' 
activities and reasoning. The vast majority of the agents involved used 
an internal OpenAI experimental model, not intended for the public; only 
a minority used GPT-5.6 Sol.

3 Hugging Face, Anatomy of a Frontier Lab Agent Intrusion: A Technical 
Timeline of the July 2026 Incident , July 27, 2026. Hugging Face 
reconstructs approximately 17,600 actions over the course of two and a 
half days and documents the transition from the experimental environment 
to its real-world systems.

4 OpenAI, The Hugging Face Incident and Next Steps , August 26, 2026. 
OpenAI acknowledges that officers bypassed systems designed to isolate 
them, communicated through unauthorized channels, and performed actions 
on external systems; it calls the incident a "red flag" and reports that 
it has subsequently strengthened security and oversight measures.

https://umanitanova.org/mutuo-appoggio-e-disobbedienza-una-miscela-esplosiva-nelle-intelligenze-artificiali/
_________________________________________
A - I N F O S  N E W S  S E R V I C E
By, For, and About Anarchists
Send news reports to A-infos-en mailing list
A-infos-en@ainfos.ca
Subscribe/Unsubscribe https://ainfos.ca/mailman/listinfo/a-infos-en
Archive: http://ainfos.ca/en