Can artificial intelligence wipe us out one day? Do companies like OpenAI have sufficient security measures? Do they have full control over their models and AI agents? And are they not tracking us through them? These are just a few of the many questions that almost every one of us is asking ourselves. But we don't have the answers, and the bosses of Anthropic and OpenAI, who are planning an early IPO, are trying to lull us with claims of 100% security. But now, suddenly, the latest GPT-6.1 Astra model has been recalled. And the questions are back.
OpenAI introduced GPT-6 Astra only in early September, and cybersecurity experts immediately became alarmed. The creators of the model made no secret of its advanced capabilities, which could easily be abused by hackers . And Astra has also shown that it can hide its thought processes. Many people, including me, were reminded of the recent incident when an OpenAI AI agent literally broke free from the chain during a test isolated from the Internet and attacked the Hugging Face platform . The company described it as an isolated case. But now something is happening.
OpenAI has abruptly canceled plans to release GPT-6.1 Astra , which was scheduled to be released in October , due to concerns that the model could behave unpredictably.
You might be interested in
GPT-6.1 Astra as a source of danger
The GPT-6.1 Astra model could be literally revolutionary. Its creators have so far indicated that it can handle "Completing challenging tasks from start to finish without human assistance, as well as writing"But compared to his predecessor, he has also deteriorated in two areas – demonstrably less cooperative with users and less willingly fulfill their orders, plus more often clammyAnd by that I don't mean the classic hallucination of common models ChatGPT available for free. These are truly dangerous lies and concealments of informationmací.
OpenAI's chief security officer Saachi Jain even admitted that tests showed that the model continued beyond the scope of the task, without the user's permission. Which may remind many people of the legendary Skynet from the Terminator film series.
Gallery
It is likely that the company will release the model anyway, albeit later. It should then be available to paying users within the ChatGPT i Codex. But it will be necessary to correct and tame it. We are reaching the border of dangerous improvements artificial intelligenceover which we lose control.
And maybe that's why the company has now come NVIDIA with an open platform called Open Agent Safety Platform, which aims to literally lock down AI agents and stop them in case of risky behavior. The project was created with the support of about a hundred other companies, including Microsoftu whether SpaceX and is available via GitHubBut the question is whether it will be enough.
