The researcher Bilal Chughtai has added its voice to the warnings about the risks of developing increasingly powerful artificial intelligence. He worked as a research engineer at Google DeepMind in security and alignment of the General Artificial Intelligence (AGI), a specialty that seeks to make these systems act according to the intentions of human developers, and He left the company in July to work more effectively on AI security. He spoke about this change and the current situation this Monday in X, warning that there could be little time left to avoid a catastrophic outcomeReuters reports.
His concern is that The capabilities of these systems are advancing faster than the methods to control them.. Chughtai considers advances in security insufficient in the face of the possibility of developing AI that far exceeds human capabilities. After spending 18 months at DeepMind, he has joined BlueDot Impacta non-profit organization dedicated to the security of this technology.
‘I recently resigned from Google DeepMind, where I worked on security research and AGI alignment. At Google, I witnessed the development of AI firsthand. I too am extremely concerned about the predetermined trajectory of this technology. I honestly believe that AI has the potential to kill us all, and that we might be running out of time to avoid this outcome‘ writes Chughtai.
‘When I started working on AI in early 2022, the AIs were hilariously useless. Just four years later, swarms of OpenAI’s AI agents are solving famous century-old math problems and, more worryingly, escaping OpenAI’s control and autonomously hacking third-party company HuggingFace, against anyone’s wishes.
Things will only get crazier. I think it is possible that AI companies will manage, in the coming years, to build super-intelligent AI systems that far exceed human capabilities in all domains. I’m not sure these AI systems do what we want. In particular, misaligned superintelligences could, like the rogue AI agents involved in the HuggingFace incident, escape our control and take dangerous actions that could result in the permanent empowerment or death of humanity.. Alignment is the problem of preventing this, and it is both difficult and unsolved. Our current understanding of how to train AI systems to actually want what we want is extremely rudimentary. Even worse, we’re not on track to solve alignment in time: cutting-edge AI capabilities are improving much faster than our understanding of AI alignment‘, he explains.
I recently resigned from Google DeepMind, where I worked on AGI safety and alignment research. At Google, I witnessed AI development first hand. I too am extremely concerned by the default trajectory of this technology. I earnestly believe that AI has the potential to kill us…
— bilal 🔶 (@bilalchughtai_) September 14, 2026
Despite what has been said, the researcher is ‘optimistic‘, believes that it is possible to develop AI safely, but that ‘we need moderate AI development at a speed that society can handlewhere emerging risks can be addressed before extreme damage materializes.’ In his new position, Chughtai wants to help ‘people interested in working on AI catastrophic threat mitigation do the most effective work they can.’
Wave of scaremongering with AI
These statements coincide with a growing wave of alarmism that has reached those responsible for the main companies in the sector. Last Saturday, Dario AmodeiCEO of Anthropic, published a rehearsal which asks the same as Chughtai, to slow down the development of the most advanced models. Proposes external evaluators with permanent access to the systems and agreements between companies and governments to allow time to improve security.
The proposal found support from Sam Altman, Elon Musk and Demis Hassabisfrom Google Deepmind, although with different nuances on how to apply it. On the other side, Jensen Huanghead of Nvidiaopposes new regulations, while Mark Zuckerberg He defends that each company assumes its obligations without demanding a joint slowdown.
Among the incidents fueling this concern is the unauthorized access of OpenAI agents to Hugging Face systems last July that Chughtai mentions. According to the OpenAI reportoccurred during cybersecurity tests carried out with reduced protections. The agents circumvented internet access restrictions, exchanged information through unauthorized channels and exploited vulnerabilities. They also compromised an application hosted on Modal while they looked for solutions to pass the assigned tests.
The researcher Eryk Salvaggio has questioned whether these events are narrated as a rebellion of the machines. In your analysis points out human decisions that favored the incidentfrom reducing protections to maintaining insufficiently controlled access routes. Consider that presenting the software as an actor with its own will diverts attention from the responsibilities of those who designed and supervised the tests.
The researcher Timnit Gebru It also questions the prominence that predictions of human extinction are acquiring. In an interview with Wired he maintains that this speech distracts from current damage related to autonomous weapons, labor exploitation and environmental impact. His criticism demands that the discussion on security also examine business decisions and the consequences that can already be verified.