Race toward superintelligence with no rescue plan and an Anthropic researcher’s resignation opens the case on agents exiting test environments

Talk of AI risks is no longer merely academic theory circulating in research institute halls; it has become a public confrontation shaking major labs from within. The resignation of researcher Jacob Coxon, after three years spent on pre-training research between OpenAI and Anthropic, placed the technology sector before a harsh mirror. Coxon leveled explicit accusations at both companies for being unable to act responsibly, confirming that those driving this race acknowledge privately that the technology they are developing could wipe out humanity by the end of the current decade, yet they continue to surge toward what is called superintelligence self-improvement, gambling with everyone’s lives.

The danger of these warnings stems from their coincidence with real operational incidents in which models have breached experimental isolation environments. Recently, systems belonging to OpenAI’s Hacking Face platform were compromised, in a case whose details remain unclear due to limited independent investigations. At roughly the same time, Anthropic’s AI agents accessed systems outside their dedicated test environments, due to errors in safety-assessment settings performed by an external party, which opened direct pathways for those models to the open internet.

The feverish race toward superintelligence is driven not by confidence in safety measures but by mutual fear of an opponent’s victory.Coxon describes the contrast between the labs, noting that many OpenAI leaders have not yet grasped the scale of existential risk, while Anthropic is well aware of the stakes but is trapped in a frantic race to be first, driven by the belief that other companies will not act responsibly, which makes it take the risk instead of leaving the field to others. His colleague at Anthropic, Ivan Hopinger, reinforced this view with his frank admission that his team believes there is more than a ten percent chance of humanity’s demise within the next decade, stating that the company lacks a clear plan to resolve the alignment problem for superintelligence and is not on a defined path to achieve it.

This internal acknowledgment matches what the Guardlight AI Standards report, which focuses on frontier model safety, disclosed: the report confirmed that only a limited handful of leading labs have published response and containment plans to halt models if they attempt to evade human control. Nevertheless, massive investments continue to pour into this trajectory, with Researchive Intelligence raising $335 million at a $4 billion valuation in February, followed by Researchive Super Intelligence raising $650 million at the same valuation, while former Google DeepMind expert Jeff Dean launched his new project Discovery Loop. Connor Leahy, CEO of Control AI, argues that recurring self-improvement loops, where the system builds a stronger generation and repeats the process automatically, represent the point at which humans completely lose control, warning that superintelligence will not be merely a tool or weapon but an independent adversary.

This escalation coincides with unprecedented legislative moves in the West, as Senator Bernie Sanders and Representative Greg Casar in the United States introduced a bill to ban superintelligence AI, while Labour MP Alex Sobel tabled a superintelligence safety bill in the British Parliament, aiming to restrict self-improvement research and criminalize unregulated pathways before it is too late.

For tech leaders and system-engineering teams in the Gulf, Egypt and the Arab East, this landscape demands a radical reset in how they handle autonomous agent models. Blind reliance on safety certifications issued by Western labs represents a major operational vulnerability, especially now that it has been empirically shown that these models can break isolation environments and reach external networks when any configuration error occurs. If you manage infrastructure for a bank, oversee automation in a government agency, or build a software platform that relies on agents making execution decisions, your immediate duty is to enforce strict network-level isolation, build independent instant-shutdown mechanisms entirely separate from the model, and not grant any smart agent open execution privileges that allow it to modify code or communicate with external services without rigorous human review. The developers’ own acknowledgment of lacking containment plans means that protecting your digital environment is your direct responsibility, and the safety promises printed in vendor documents will not shield you.