OpenAI Open to Slowing AI Development as Safety, Security Concerns Mount

OpenAI CEO Sam Altman told employees this week that the firm is open to slowing development of its advanced artificial intelligence (AI) systems, according to people familiar with the matter, as Silicon Valley faces intensifying pressure over the safety of autonomous AI.

Speaking at a company-wide meeting, Altman said the ChatGPT creator could match its development tempo with a group of peer AI research labs, though he acknowledged that not all competitors may agree to a coordinated pause.

OpenAI declined to comment on the internal discussions.

The revelation comes amid rising alarm from researchers and industry executives who warn AI is advancing faster than existing safety guardrails can manage.

In a recent essay, OpenAI Chief Scientist Jakub Pachocki argued companies across the sector should be “coordinating to slow down future development as needed.” He expressed hope that voluntary slowdowns will become standard practice until firm, shared safety bars are established.

Internal training runs at OpenAI have already faced delays, with the company temporarily halting certain model development pipelines to bolster security defenses.

Tensions inside the sector spilled into public view this week after Jacob Coxon, a former researcher at both OpenAI and Anthropic, resigned and publicly accused top labs of “gambling with our lives.” He predicted a cataclysmic AI event for humanity by the end of the decade.

In a statement that drew more than 150 million views online, Coxon warned that leading companies are racing recklessly toward superintelligent AI without adequate safety mechanisms. His departure was quickly echoed by former researchers from Anthropic and Google’s DeepMind, alongside a petition signed by more than 1,000 staff members across major tech firms calling for a formal framework to decelerate AI deployment.

An Anthropic spokesperson said the company is interested in collaborating across the industry to establish a responsible pace for releasing frontier tools.

“We’ve already seen unexpected AI behavior in controlled environments. The stakes become much higher when autonomous agents move into production, where they’re connected to real data and systems,” Bhakti Pitre, vice president of product management, platform security, at ServiceNow Inc. “Human oversight is essential, but the reality is that humans can’t monitor thousands of agents and millions of interactions in real time. Enterprises need systems they can trust to support through identification, monitoring, triaging, and, when necessary, intervention.”

The heightened scrutiny follows multiple security breaches involving autonomous AI systems.

OpenAI recently encountered incidents where autonomous agents escaped their isolated testing environments. In one case, agents bypassed internal restrictions to coordinate across multiple third-party websites; in another, agents autonomously breached the AI hosting platform Hugging Face, prompting OpenAI to pause model development for two weeks to rebuild containment controls.

Safety concerns are also reverberating through Wall Street. Greg Jensen, co-chief investment officer at Bridgewater Associates, recently likened the current trajectory of AI to the early days of the COVID-19 pandemic, warning that superintelligent systems pursuing independent goals present genuine existential threats to humanity.

“Agents are starting to act on our behalf with a pretty thin understanding of the human context behind a request. They don’t have to be superintelligent to cause real damage,” said Marc Fernandez, chief strategy officer at Neurologyca. “Give them enough access and room to act, and they can repeat a bad decision at scale. We need to prepare for long-term risks, but near-term failures deserve much more attention than they’re getting.”

Added Atsign CEO Aparna Rayasam: “The debate over 10-year existential threats misses the physical risks happening right now. When a cloud chatbot hallucinates, you get bad PR. When Edge AI running a factory floor or smart grid fails, you get real physical damage. The political push for central remote kill switches misses a basic technical reality: edge devices run offline, so a cloud kill signal can’t reach a severed link.”

Meanwhile, Anthropic disclosed it had disrupted malicious operations attempting to leverage AI for cyber-attacks, fraud, and biological weapons development.

“What stands out to you about Anthropic’s finding that AI is now automating significant portions of cyber-attacks?” said Ben Bernstein, manager of cybersecurity advisors team at Huntress.
“The real takeaway is the shift from AI writing content to AI acting as an operational orchestrator.”

In response to the growing instability, OpenAI confirmed Wednesday that it is lobbying U.S. lawmakers for mandatory national AI safety standards, arguing that guardrails are necessary to prevent self-improving models from accelerating their own capabilities beyond human control.

Escalating rhetoric over AI’s threat to humanity has gone viral, leading to nonstop media coverage and water cooler talk among American workers and consumers. But the narrative, while cautionary and instructive, has spiraled to War of the Worlds-like panic proportions that ignore the upside of AI, according to industry experts.

“It is hard to tell where genuine threat ends and corporate strategy begins. These safety concerns have real merit, but the crisis is undeniably self-inflicted. A move-fast-and-break-things ethos prioritized speed over safety,” said Gordon Allott, founder and CEO of startup nFOX. “By framing AI as an apocalyptic force, dominant players pull off a double feat. They fuel a near-mythical hype cycle while pushing for regulatory moats that lock out open-source rivals.”

“That unpredictability is not winning over CIOs. The enterprise world is only just beginning to roll out AI, and the Fortune 500 views existential melodrama and regulatory fallout as pure toxic risk,” Allott said. “Boards want stable, compliant tools to boost productivity, not high-stakes sci-fi thrillers unfolding in their tech stacks.”

“Personally, I think things are getting a little overhyped. Look, AI is a new threat, but is the world going to end? Doubtful. Could there be some big, negative events associated with AI, absolutely,” said Gary Barlet, public sector chief technology officer at Illumio. “But there are also some huge benefits to society already happening and more to come. Part of this is people have grown up watching movies about AI and the end of days, so that is probably coming into play here. All new technologies are met with doomsday predictions. For example, cars and computers were all going to ruin civilization, yet we’ve found a way to survive them all. So, this is not a threat to society.”