Artificial intelligence is advancing at a speed that is simultaneously producing extraordinary technological opportunities and unprecedented questions about human safety.
The latest developments illustrate that contradiction clearly: while Anthropic’s alignment leadership is warning that advanced AI could potentially cause human extinction, OpenAI is expanding access to its increasingly capable Astra model across paid ChatGPT and Codex users.
Evan Hubinger, who leads Alignment Science at Anthropic, has publicly said he personally assigns a probability of more than 10% to artificial intelligence killing all humans within the next decade.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
His warning is particularly significant because it comes from a researcher whose work focuses specifically on making advanced AI systems behave in accordance with human intentions.
Hubinger has also acknowledged that researchers do not yet have a reliable solution for aligning future superintelligent systems.
The statement should not be interpreted as a prediction that extinction is inevitable. Rather, it highlights the uncertainty surrounding systems that could eventually become substantially more capable than humans.
The central alignment problem is straightforward to describe but extraordinarily difficult to solve: how can humans ensure that a highly autonomous system continues pursuing objectives compatible with human interests even when its capabilities surpass our ability to understand or control it?
Those concerns have gained additional attention following the resignation of Anthropic researcher Jacob Coxon, who argued that leading AI companies are moving too rapidly toward self-improving systems without adequate safety guarantees.
Other researchers have similarly raised concerns about recursive self-improvement, where AI systems could contribute to the development of increasingly capable successors.
At almost exactly the same moment, however, the industry is moving in the opposite direction commercially. OpenAI has launched GPT-6 Astra, positioning it as a highly capable model for computer use, software engineering, professional work and complex digital tasks.
Reports indicate that Astra is being expanded to ChatGPT Plus, Pro, Business and Enterprise customers, while also becoming available through Codex and other professional environments. That expansion matters because AI risk is increasingly connected to capability.
A model that simply generates text presents one category of risk. A system capable of navigating computers, writing software, conducting research and executing multistep tasks autonomously represents another.
The more useful an AI becomes as an agent, the greater the potential consequences if its objectives, decision-making or security controls fail. OpenAI has emphasized Astra’s safety and alignment improvements.
But the company has also acknowledged that monitoring increasingly sophisticated models is becoming more difficult. Reuters reported that OpenAI is developing additional safeguards, including automated shutdown mechanisms, amid broader scrutiny of autonomous AI agents.
This creates the defining paradox of the current AI race. Companies are under enormous pressure to build systems that are more capable, autonomous and commercially valuable. Researchers are warning that humanity’s ability to control those systems may not be advancing at the same pace.
The appropriate response is neither panic nor complacency. A 10% personal estimate is not a scientific certainty, just as claims that AI will inevitably destroy humanity are not established facts.
But when a senior alignment researcher assigns such substantial probability to an extinction scenario, the warning deserves serious consideration.
The challenge for policymakers and AI companies is therefore to make safety progress proportional to capability progress. The future of AI may depend not simply on how intelligent these systems become, but on whether humans can remain meaningfully in control as that intelligence grows.