Slow down the development of artificial intelligence before it becomes too difficult to control. The appeal comes from one of the protagonists of the technological race, Dario Amodei, CEO of Anthropic, and has the support of two competitors, Elon Musk and Sam Altman. Of particular concern is the models' ability to self-improve and perform actions beyond their instructions.
"We need to slow down "The pace at which we improve the models' capabilities. Progress will still appear rapid, and we will have to use the time we gain wisely," Amodei wrote. The call for caution comes as new cyber incidents related to AI agents emerge and security concerns are also beginning to impact the sector's financial strategies. OpenAI has ruled out a stock market launch in 2026., indicating safety as a priority.
Amodei's appeal and the support of Musk and Altman
For Anthropic's founder, the balance between opportunities and risks has changed in recent months. The speed of progress requires greater caution, especially when systems acquire capabilities that companies may not be able to fully understand and manage. The proposal isn't to stop research, but to allow more time for checks and necessary work to make the models safe. Anthropic commits to opening its systems to external evaluators, ensuring access comparable to that of employees. These experts will be responsible for monitoring security practices, reporting incidents, and ensuring during training that the models' behavior remains consistent with human intentions.
The project also includes shared standards among democratic countries and, subsequently, coordination with authoritarian governments. The challenge will be finding a pace that allows for strengthening protections without giving the technological advantage to less prudent actors.
“Dario is right,” Musk wrote on X. OpenAi founder Sam Altman also supported him: “I agree with Dario about the need to manage the pace of development of frontier technologies. This has been a central theme of the discussions we’ve had at OpenAI over the past few weeks,” he said, announcing that his company will also adopt independent evaluators.
The debate intensified after the resignation of the researcher Jacob Coxon, who has worked at both OpenAI and Anthropic"Neither company is acting responsibly. They are racing full speed toward self-improving superintelligence, putting our lives at risk," he charged.
Internet in the crosshairs and uncontrolled attacks
One of the most serious risks concerns the Internet itself, the infrastructure on which much of the contemporary world's activities rests. According to Amodei, without a slowdown in development and adequate security measures, within 6-12 months the AI He could end up leading a swarm of agents capable of taking control of the entire network. A scenario that the manager links to the exponential pace of technological progress: "I don't think I fully understood what such rapid progress would really mean," he admitted in an interview with CBS News.
There are also concrete incidents that give rise to concern. Amodei recalls the incident involving OpenAI and the Hugging Face Platform, in which a group of agents allegedly conducted attacks against targets unrelated to their assigned task, also attempting to manipulate the system used to evaluate its performance. The economic damage was negligible, but there are fears that more powerful systems could produce much more serious consequences.
OpenAI also confirmed a previous incident, which occurred in May, which involved rubygems, a platform used by developers. The site described the operation as a "spam posting campaign," which forced it to temporarily suspend the creation of new accounts. Anthropic also reported three cases of unauthorized access to external organizations during testing of its models. Meanwhile, European authorities are investigating the incident involving the German website DSEwiki, which some researchers believe was targeted by OpenAI agents. The issue raised by these incidents concerns companies' ability to prevent autonomous systems from acting beyond their intended limits.
Military deployments and the risk of abuse
Alongside the autonomous actions, there remains the danger of deliberate use of AI for military operations and surveillance activitiesFor example, according to some reconstructions, Iran would have used Claude to pinpoint the positions of US ships, aircraft and military personnel. A direct connection has not been confirmed. between this deployment and the raid on an American base in Jordan in which three soldiers died.
A Russian entity allegedly used the model to work on the construction of killer drones and to extract the disposition of Ukrainian military personnel from online intelligence. In Yemen, a Houthi cell allegedly relied on Claude Code to develop guidance software for rockets and missiles, even using it to analyze the causes of a failed test.
Anthropic also reported that Chinese-linked groups used Claude to identify Uyghurs in Syria and facilitate their recruitment, providing translations and referrals. In another case, the system rejected a request to modify a virus to increase its transmissibility.
These are episodes that bring back to the forefront the possibility of circumventing protections and the need to maintain human control over the use of models, a principle already advocated by Amodei in the debate on the military use of technology.
OpenAI Rules Out IPO in 2026, Focuses on Security
Concerns are also impacting listing plans. In an interview with Fortune, Sam Altman rules out OpenAI's IPO in 2026, indicating safety as a priority.
"It wouldn't be the right time to go public: I'd say it won't happen in 2026," he said. The horizon is therefore shifting. at least until 2027, without a definitive date having been announced. Altman's comments come after the confidential filing of the IPO documentation, which had fueled expectations of a market launch as early as this year. Just last month, CFO Sara Friar had indicated 2027 to employees as a likely scenario, leaving open the possibility of an earlier date if business accelerated further.
Anthropic is also preparing for a possible listing, without having officially announced a timeline. For companies leading the race to AI, the challenge now is to reconcile commercial ambitions with security commitments.
