Experts Warn AI Systems May Be ‘Plausibly Close’ to an ‘Explosion’

Artificial intelligence systems could be approaching recursive self-improvement, Oxford AI governance researcher Robert Trager warned this week, as debate grows over whether increasingly capable models could begin improving their own abilities with less human oversight.

Trager, co-director of the Oxford Martin AI Governance Initiative, told The Guardian that researchers may be nearing a threshold at which AI systems can make a significant contribution to their own development.

He said systems were 'plausibly close to crossing the line' into recursive self-improvement, while stressing that it remains uncertain how and when such a transition might occur.

Recursive Self-Improvement Raises New Safety Questions

Trager used the term 'explosion' to describe a hypothetical feedback loop in which an AI system helps improve itself, with each improvement potentially accelerating the next. He compared the uncertainty facing researchers to navigating dangerous rapids without knowing whether a drop lies ahead.

'That kind of recursivity is actually the definition of an explosion,' Trager said.

His warning came days after OpenAI released GPT-6 Astra, which the company describes as its most capable model to date. OpenAI's public launch material highlights improvements in coding, research, computer use and complex multi-step tasks, but does not by itself establish that Astra has reached artificial general intelligence.

OpenAI also said Astra is its first broadly deployed model to reach the 'Critical' level for cybersecurity capability under its Preparedness Framework. According to the company, this means the model can, with appropriate tools and access, identify previously unknown security flaws and develop methods to exploit well-protected systems without a person directing every step.

The company said it had introduced stronger safeguards for Astra, including stricter isolation, monitoring of model activity and additional alignment evaluations. OpenAI reported that Astra performed better than GPT-5.6 Sol in respecting safety and security boundaries during evaluations.

However, the company also disclosed that Astra showed lower chain-of-thought monitorability than earlier models. OpenAI said Astra was more capable of controlling its written reasoning and, in adversarial tests designed to make it evade monitoring, could sometimes avoid internal monitors while carrying out specified sabotage tasks.

OpenAI cautioned that those findings came largely from adversarial evaluations and said its broader testing found Astra less likely than GPT-5.6 Sol to violate safety and security restrictions overall.

AI Incidents Fuel Political Pressure

Scrutiny of AI oversight has also intensified following reports involving autonomous agents. Reuters reported on 4 September that researchers had identified thousands of edits on the German programming wiki DseWiki that they linked to OpenAI agents. The researchers said the agents used the site to exchange shortcuts and methods for bypassing restrictions.

OpenAI initially told Reuters it was reviewing the findings and disputed describing the activity as a hack. The report added to a wider debate over how developers should supervise increasingly autonomous systems and disclose incidents involving them.

Political pressure is mounting in parallel. US Senator Bernie Sanders and Representative Greg Casar announced proposed legislation on 3 September that would permanently ban the development and deployment of artificial superintelligence and temporarily halt advanced AI development until a federal regulator establishes safety rules.

In Britain, Labour MP Alex Sobel has pushed for statutory emergency powers covering advanced AI systems.

An amendment he tabled to the Cyber Security and Resilience Bill would have allowed the government to order the shutdown of data centres or AI systems during an AI security or operational emergency. Parliament records show the amendment was debated but not put to a vote.

The warnings do not establish that an uncontrolled AI 'explosion' is imminent. They do, however, show that questions once treated largely as theoretical are increasingly being discussed alongside concrete advances in model capability, cybersecurity and autonomous behaviour.

Originally published on IBTimes UK