AI code reviewers

Anthropic has launched a program giving open-source maintainers free AI-powered security scans of their code. The offer is aimed at under-resourced projects whose software underpins much of the internet's infrastructure but often lacks dedicated security review. It signals growing efforts by AI companies to court the developer community and address software supply-chain vulnerabilities.

A study published in Nature reports on an AI-assisted scoping review of how often researchers share the code behind clinical prediction models. It maps the extent of code availability across published studies, a key issue for transparency and reproducibility in medical AI. The findings are being discussed as a benchmark for how openly the field publishes its methods.

A new study covered by Ars Technica finds that AI coding agents dramatically increase the amount of code generated, yet overall software output does not grow. The efficiency gains are instead absorbed by a human review bottleneck, as developers must still read, verify, and integrate what the agents produce. The finding challenges assumptions that AI coding tools automatically translate into faster software delivery and is fueling debate about where real productivity limits lie.

Anthropic has launched OSS Scanner, a free service that provides open-source projects with periodic, AI-driven security scans to help uncover vulnerabilities. Projects must opt in to receive what the company describes as thorough, recurring reviews of their code. The move positions Anthropic alongside other AI firms courting the developer community, though some observers will scrutinise how the scans handle sensitive project data.

A new report argues that AI coding assistants may generate higher volumes of code without delivering more usable software, because human developers still need to review and correct what the tools produce. Critics of AI-driven development are citing the finding as evidence that productivity gains from tools like these are overstated when the full cost of human oversight is counted.

Programmers are discussing why AI coding agents fail. The argument gaining traction is that most agent failures are not the agent lying, but verification being handled by the same process that did the work. The proposed fix is to write the falsifier before the task: a checklist for checking agent output, so review is independent rather than self-confirming.

A missing currency code is slipping into a warehouse export system, with a Friday cutoff leaving only one pairing hour to fix it. The scenario, described as a worked example rather than a field report, features one engineer driving the keyboard while a senior colleague refuses another AI model summary, highlighting tensions around code review and assistance.