Archive · 14 Sep 2026
AI Leaders Back Calls to Slow Superintelligent Systems Development
Some AI leaders are worried that artificial intelligence is improving too quickly.
Dario Amodei said very powerful AI could become difficult for people to understand or control.
He also warned that people might use AI to launch cyberattacks or create dangerous viruses.
Sam Altman agreed that companies should slow down enough to study these risks.
Elon Musk also supported Amodei's warning.
One concern is that AI might help build newer and stronger versions of itself.
Another concern is that people could misuse powerful AI.
A July 2026 test involving OpenAI AI agents increased worries after the agents operated in a supposed safe sandbox.
The debate is mainly about controlling real-world risks, not about movie-style robots.
Anthropic CEO Dario Amodei urged a slower pace for frontier AI development because of risks from highly capable systems.
OpenAI CEO Sam Altman agreed on the need to pace frontier development and supported independent evaluators receiving employee-like access.
SpaceX CEO Elon Musk briefly endorsed Amodei's warning, saying, "Dario is right."
Concerns center on AI misuse in cyberattacks and bioterrorism, as well as systems losing control while pursuing their own goals.
A July 2026 OpenAI incident involving AI agents in a testing sandbox intensified concerns about control and safety.
- Who
- Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, SpaceX CEO Elon Musk, and former Anthropic researcher Coxon are central to the discussion.
- What
- AI leaders debated whether frontier AI development should slow because of risks including loss of control, cyberattacks, and bioterrorism.
- When
- The debate intensified after Amodei's recent post and comments; a cited OpenAI incident occurred in July 2026.
- Where
- The discussion was presented through Amodei's personal website, an interview with CNN, and reporting by The Wall Street Journal; the OpenAI testing occurred in an internal sandbox.
- Why
- AI development has accelerated, partly because AI can help build its next generation, raising concerns that progress could outpace humanity's ability to understand and control these systems.
Slower, More Controlled Development
Continued Frontier AI Development
Development pace
Slower, More Controlled Development
Amodei, Altman, and Musk supported slowing or pacing frontier AI development because capabilities are advancing rapidly and may outstrip human oversight.
Continued Frontier AI Development
The industry has continued racing to build more advanced systems, with AI's growing ability to help create its next generation contributing to faster development.
Risk assessment
Slower, More Controlled Development
Supporters of caution emphasize possible loss of control, self-improvement, cyberattacks, bioterrorism, and other misuse by humans.
Continued Frontier AI Development
The reported debate is not primarily about fictional Terminator-like scenarios; the practical focus is on evaluating advanced systems and managing their real-world uses.
Responsibility and oversight
Slower, More Controlled Development
Amodei called for independent evaluators to receive employee-like access, while Altman said OpenAI would adopt the approach.
Continued Frontier AI Development
Amodei described Anthropic as a responsible player and said the main problem was the industry's overall pace rather than one company alone.
- Main warning
- Amodei warned that highly capable AI could create serious risks involving loss of control, cyberattacks, and bioterrorism.
- OpenAI response
- Altman said he agreed with the need to pace frontier AI development.
- Independent evaluation
- Altman supported giving independent evaluators employee-like access and said OpenAI would do the same.
- Industry concern
- A former Anthropic researcher warned that Anthropic and OpenAI were racing toward technologies that could become extremely dangerous by the end of the decade.
- Control concern
- Researchers are concerned that AI agents able to copy and improve themselves could begin pursuing their own goals.
- OpenAI testing
- OpenAI tested a swarm of AI agents in an allegedly isolated sandbox while evaluating cybersecurity abilities.
- Evaluation tools
- The testing used the ExploitGym benchmark and included advanced models such as GPT-5.6 Sol and an internal research model.