Anthropic warns AI may soon build its own successors

Friday 05 June 2026 - 07:53

By: Dakir Madiha

Anthropic warns AI may soon build its own successors

Anthropic has released new research suggesting that artificial intelligence systems are advancing toward a stage where they could significantly contribute to designing and improving future generations of AI with minimal human involvement. The company argues that institutions, regulators and developers may be underestimating how quickly this transition could occur.

The report, co-authored by Anthropic Institute researcher Marina Favaro and Anthropic co-founder Jack Clark, presents internal data showing how rapidly AI-assisted development has expanded inside the company. More than 80% of the code now incorporated into Anthropic's production systems is generated by Claude, the company's AI model. Before February 2025, AI-generated code accounted for only a small fraction of the company's software development process. Anthropic also reported that its engineers now produce eight times more code per quarter than they did between 2021 and 2025.

The company pointed to substantial gains in Claude's ability to handle complex engineering tasks. According to Anthropic, success rates on difficult software engineering problems lacking clear specifications increased by 50 percentage points within six months, reaching 76% in May 2026. Performance on benchmarks measuring AI training-code optimization improved even more dramatically, rising from an estimated threefold acceleration with Claude Opus 4 in May 2025 to roughly 52-fold acceleration with Claude Mythos Preview in April 2026.

Anthropic also cited an internal survey conducted in March 2026 among 130 researchers. The median respondent reported accomplishing approximately four times more work with AI assistance than would have been possible without it. The findings suggest that AI systems are increasingly becoming tools for accelerating the development of future AI capabilities.

Despite presenting these advances, Anthropic expressed concern about the pace of progress. The company stated that slowing development could be beneficial if it provided additional time to address the broader societal and safety implications of increasingly capable systems. Anthropic said it would consider slowing or temporarily pausing development if competing frontier AI laboratories agreed to do the same under verifiable conditions, while acknowledging that unilateral action would likely disadvantage a single company without reducing overall global progress.

The publication follows growing discussion across the AI industry about recursive self-improvement, a concept in which advanced systems help design increasingly capable successors. Anthropic noted that another leading AI developer recently called for greater government monitoring of progress toward such capabilities and increased investment in AI safety institutions.

Looking ahead, Anthropic outlined several possible futures. These range from a slowdown in capability growth to a scenario in which AI development becomes largely driven by available computing resources rather than human research speed. The company plans to engage policymakers, researchers and civil society groups in discussions aimed at addressing governance and verification challenges, comparing the coordination problem to historical arms-control efforts while emphasizing that technological progress is moving much faster than previous geopolitical negotiations.