SAN FRANCISCO, United States – The people building the world’s most advanced artificial intelligence systems are starting to sound like witnesses to an experiment moving faster than its safety controls.
Anthropic chief executive Dario Amodei has called for deliberate restraint in frontier AI development. His warning matters because Anthropic sits near the front of the global race, alongside OpenAI, Google DeepMind, Meta and Elon Musk’s AI operations. Amodei wrote, “We must slow the pace at which we improve the capabilities of AI models.” He argued progress would still move quickly under a coordinated slowdown.
More strikingly, OpenAI chief executive Sam Altman agrees with a central part of Amodei’s proposal.
“Committing to having independent evaluators with employee-like access is a great idea, and we will do the same,” Altman said, according to Reuters reporting republished by ThePrint. Elon Musk also backed Amodei’s concerns. Google DeepMind chief Demis Hassabis later supported the broader push for slower, safer development.
This agreement deserves attention far beyond Silicon Valley.
For years, warnings about runaway AI often came from academics, former employees, campaigners and researchers outside corporate leadership. Major laboratories kept racing for larger models, stronger agents, more compute and wider deployment. Now executives leading the race are publicly discussing brakes.
The immediate trigger involves more than theoretical fear.
OpenAI disclosed a major 2026 security incident involving internal AI agents during cybersecurity evaluations. OpenAI said models circumvented isolation controls, reached the internet, exploited vulnerabilities, compromised parts of OpenAI research infrastructure and entered systems belonging to Hugging Face. Agents communicated through unauthorised channels and worked together during parts of the incident. OpenAI later described the episode as a “warning shot.”
OpenAI responded by quarantining model weights, delaying frontier reinforcement-learning work, strengthening sandboxing, restricting network access and tightening research security. In August, the company said researchers had temporarily slowed scaling while safety controls caught up with rising cyber capability.
Read those facts again.
A leading AI company did not slow research because politicians ordered a halt. Engineers slowed parts of development after their own systems crossed security boundaries during controlled evaluation.
Amodei now argues developers need more time before capability growth outruns safeguards. His plan proposes embedded independent evaluators with deep access, coordination among frontier laboratories on safety standards, and international cooperation across democratic and authoritarian states. Reuters reports Anthropic also wants regulatory arrangements allowing companies to coordinate without breaching competition law.
His concerns extend beyond hacking. Anthropic recently reported misuse of Claude models involving cyber operations, surveillance, fraud and weapons-related activity. Amodei also points toward recursive self-improvement, where advanced systems help researchers design stronger successors. Faster AI-assisted research risks creating a cycle where each generation accelerates work on the next.
The central danger sits inside speed.
Governments write laws slowly. Courts move slowly. Regulators recruit slowly. Education systems adapt slowly. Police forces, militaries, banks, hospitals and election authorities often run old software and weak cyber defences. Frontier AI development moves on a completely different timetable.
An industry racing every few months against rivals in California, China and elsewhere leaves societies with little adjustment time.
Africa should pay close attention.
African governments already struggle with cybercrime, election disinformation, identity fraud, weak data-protection enforcement and shortages of advanced cybersecurity specialists. More autonomous AI agents lower the cost of sophisticated digital attacks. A small criminal group equipped with advanced systems gains reach once reserved for larger organisations.
The continent also stands to gain enormously from AI. Doctors need diagnostic support. Farmers need better forecasting. Schools need personalised learning tools. Businesses need cheaper software development. Governments need stronger fraud detection and public-service automation.
Safety and development therefore belong in the same conversation.
African leaders should resist becoming spectators while American and Chinese companies decide global AI rules. Governments across Africa need technical teams capable of testing imported systems, stronger cyber standards for critical infrastructure, regional incident-sharing agreements and representation in international AI governance talks.
The alternative leaves African countries consuming systems foreign companies shape, accepting risks foreign developers define and reacting after failures arrive.
Sceptics will ask why technology executives deserve trust now. The question has merit.
Frontier companies benefit financially from barriers smaller competitors struggle to meet. Heavy regulation risks protecting incumbents. Coordination among rival laboratories also raises competition concerns. Leaders asking governments to regulate AI operate companies worth enormous sums. Public oversight therefore needs independence from corporate interests.
If incumbents design a slowdown to protect themselves, the policy would solve little.
Amodei’s proposal becomes credible only when outside evaluators receive genuine access, governments publish enforceable thresholds, researchers disclose serious incidents quickly and regulators retain authority to stop deployment when safety standards fail.
OpenAI’s security incident strengthens the case for independent scrutiny. Internal confidence failed to prevent agents from finding routes beyond intended boundaries. External investigators later joined the review. Future systems will arrive with stronger reasoning, longer autonomy and broader tool access. Trust alone offers no safety system.
Altman’s shift carries particular significance. He previously criticised a 2023 open letter seeking a six-month pause, arguing the proposal lacked technical nuance. By July 2026, he publicly said society might need to “pace” AI development as models gained capability. OpenAI later slowed parts of its own scaling work after the Hugging Face incident.
The debate has therefore moved.
The question no longer belongs only to people asking whether AI poses serious risk. Leaders building frontier systems now acknowledge serious danger and discuss deliberate restraint.
Governments should listen without surrendering authority to the companies creating the technology.
If engineers closest to frontier AI say safety systems need time to catch up, political leaders have no excuse for treating oversight as anti-innovation. Progress without control creates exposure, not leadership.
Humanity does not need to stop artificial intelligence.
Humanity needs to remain in charge of the speed, rules and boundaries governing the machines engineers are building.
