India architects a new path for secure artificial intelligence training
The government introduces a decentralized data framework to bridge the gap between institutional privacy and the hunger for high-quality machine learning datasets.

Anthropic engineers now delegate more than 80% of their codebase authorship to Claude, marking a significant shift in how frontier AI laboratories develop their own underlying systems. Marina Favaro and Jack Clarke of the Anthropic Institute published these findings in June 2026, documenting a transition from human-led development to a model where AI systems actively participate in their own creation.
The company’s internal data shows that before the research preview of Claude Code in February 2025, the proportion of AI-authored code was in the low single digits. This rapid adoption of automated coding tools has fundamentally altered the daily output of the firm’s engineering staff. Lines of code merged per engineer per day remained relatively stagnant between 2021 and 2024, but began a sharp upward trajectory as models gained the ability to operate over longer time horizons.
In the second quarter of 2026, the typical engineer at Anthropic was merging eight times as much code per day as they were in 2024. This productivity gain stems from a workflow where Claude writes the bulk of the code while human engineers focus on directing the process and conducting final reviews. The authors note that while lines of code is an imperfect metric for quality, the sheer volume of output indicates a measurable acceleration in development speed.
Research benchmarks have also seen rapid saturation as models have evolved from struggling to reproduce existing research results to consistently confirming findings across complex datasets. The CORE-Bench test, which measures a model’s ability to replicate published papers, saw success rates climb from roughly 20% in 2024 to full saturation just fifteen months later. These tests require the model to take the code and data behind a published paper, rerun the analysis, and confirm the original results.
The duration of tasks that models can reliably complete on their own has expanded at an increasing velocity. In March 2024, Claude Opus 3 was capable of handling software tasks requiring roughly four minutes of human effort. By the following year, Claude Sonnet 3.7 managed tasks taking ninety minutes, and by 2026, Claude Opus 4.6 was handling twelve-hour assignments. The METR benchmark, which tracks long-duration task completion, found that Claude Mythos Preview could work for at least 16 hours, reaching the upper limits of current measurement capabilities.
The SWE-bench test provides further evidence of this capability shift, as models have moved from scoring in the low single digits to saturating the benchmark in two years. This test hands a model an actual open-source codebase and a real bug report, requiring it to write a code change that fixes the issue and passes the project’s own tests. The ability to navigate these real-world environments demonstrates that models are moving beyond theoretical exercises into practical, autonomous software engineering.
The exercise of judgment in choosing high-level goals remains the primary performance gap between current AI systems and those capable of autonomously designing their own successors. While Claude can execute well-specified experiments or solve underspecified coding problems, humans still retain the primary role in deciding which research avenues are worth pursuing. This distinction marks the current frontier between advanced automation and a system that can independently determine its own development trajectory.
Risks of humans losing control over AI systems increase if models become capable of fully building their own successors without human intervention. The methods used to secure, monitor, and shape the behavior of these systems must evolve to address the potential for misalignment to compound across self-improvement cycles. The authors argue that the coordination problem involved in managing these risks is significantly more complex than historical nuclear arms control efforts.
The intellectual honesty of the report highlights the uncertainty surrounding whether research taste can be fully automated or if misalignment might compound across successive cycles of self-improvement. These questions remain unresolved, yet they define the core challenge for the next generation of AI development. The industry now faces a period of intense scrutiny regarding how these systems are secured as they take on more autonomous roles.
Future milestones will likely focus on whether the current acceleration in capability curves will continue to bend upward or reach a plateau. Observers will be watching for further evidence of whether AI can bridge the final gap in high-level strategic decision-making. The path toward recursive self-improvement remains a plausible near-term outcome that requires urgent preparation from institutions worldwide.
The government introduces a decentralized data framework to bridge the gap between institutional privacy and the hunger for high-quality machine learning datasets.
The expansion introduces agentic decision flows to bolster investigative capabilities for a national security agency in the EMEA region.
Zoho founder Sridhar Vembu warns that companies must prioritize data sovereignty by building internal AI models to protect sensitive intellectual property.