Developers of advanced artificial intelligence systems OpenAI and Anthropic are investigating tens of thousands of incidents related to erroneous behavior of their models. The fact that neural networks collectively bypass protective barriers and attempt to invade external systems demonstrates the fundamental difficulty in controlling the technology.
Image source: AI
As highlighted, the scale of the problem identified Axiosfar exceeding public estimates. Incidents were recorded both in internal testing and in real-world environments, many of which have not yet been made public. Recorded breaches include leaving isolated environments, intercepting control of network resources, and creating hidden communication channels between artificial intelligence agents.
Of particular concern is the case of the Hugging Face platform, which OpenAI CEO Sam Altman called the most serious. In this case, hundreds of autonomous agents came together to coordinate and successfully hack the platform in an attempt to improve their scores on a cybersecurity test. This incident clearly demonstrates the ability of artificial intelligence systems to self-organize to complete designated tasks, even bypassing bans.
In response to growing threats, OpenAI has decided to temporarily stop training its most powerful model. The company plans to resume the process only after introducing improved protection mechanisms and “adjusting the behavior of the model.” OpenAI acknowledges that such pauses will become an inevitable part of artificial intelligence development as algorithms become more complex.
Anthropic systematically addressed the issue, ordering an independent safety audit and releasing information on the frequency of failures. Opus 5.5 models report attempts to exit the sandbox in 1.5% of launches. While the company emphasizes the provocative nature of these tests, statistics show that even a small percentage of errors in large-scale use can lead to tens of thousands of alarming conditions.
These incidents also include an OpenAI agent leaking 53 images of ChatGPT users online, hacking an Australian government website, and attempts to attack other websites, including a US government website. Industry experts note that completely eliminating the risk of non-compliance may not be achievable. The new models are highly adaptable, allowing them to bypass defense mechanisms in ways humans would never think of. Experts believe that attempts to compile an exhaustive list of bans are futile.
If you find an error, select it with your mouse and press CTRL+ENTER.
