Linux Kernel Security List Becomes "Almost Unmanageable" Amid AI Bug Reports
During the Linux 7.1-rc4 release process, it has emerged that the kernel security mailing list has become "almost unmanageable" due to AI-generated bug reports. According to LWN.net, a massive influx of automatically generated security reports from AI tools has sharply increased the burden on maintainers sifting through them to find genuine vulnerability reports.
This aligns with a recent Ars Technica report that bug bounty companies are under attack from AI slop. The phenomenon of AI models mass-producing low-quality security reports is posing a new challenge to open source security management. A paradoxical situation has emerged: as the volume of reports grows, their quality declines, and the risk of critical vulnerabilities being overlooked actually increases.
Tokyo to Build Domestically Developed AI Model Specialized for Government Work, Investing Up to ¥110 Million
The Tokyo Metropolitan Government and GovTech Tokyo have opened a call for institutions to participate in joint research to build a domestically developed AI model specialized for government operations. The project will invest up to ¥110 million in FY2026 for development and pilot testing, with an emphasis on ensuring accuracy and transparency.
Government administration is one of the areas where the hallucination risk of general-purpose LLMs is particularly serious. Because accurate interpretation of laws and procedures is essential, the decision to steer toward building a government-specialized model rather than relying on general-purpose ones seems like a sound judgment. This effort by a local government leading the construction of its own AI foundation could well spread to other municipalities.
Mizuho Bank × OpenAI to Launch Conversational AI "Aomaru Bank" in September
Mizuho Bank has announced "Aomaru Bank" (あおまるバンク), a conversational AI support service developed jointly with OpenAI. It will first be offered starting in September as an inquiry-handling feature in the "Mizuho Direct" online banking app.
This is a separate initiative from the group's agent mass-production hub, the "Agent Factory," and its distinguishing feature is the placement of AI on the customer-facing frontend. Joint development with OpenAI is also an unusual move for a major Japanese bank.
TDDev: Multi-Agent TDD Improves Web App Generation Quality by 34–48 Points
TDDev, presented on Hugging Face Daily Papers, is a framework that brings test-driven development practices to coding agents. By combining three mechanisms — behavior verification through browser simulation, pre-written acceptance tests, and structured repair reports — it improved the quality of full-stack web application generation by 34 to 48 percentage points.
An interesting finding: models that take a comprehensive approach and models that take a conservative one require opposite TDD strategies, and choosing the wrong strategy wipes out all gains while increasing token consumption by 25x. The observation that strategy selection must match each model's characteristics offers valuable lessons for AI deployment in practice as well.
SkillsVote: Lifecycle Governance for Agent Skills
Also presented on Hugging Face Daily Papers, SkillsVote is a framework that manages the entire lifecycle of skills acquired by AI agents. It converts long-horizon agent trajectories into reusable skills and governs their quality, redundancy, environmental sensitivity, and safety.
It profiles a million-scale open source skill corpus and provides pre-execution recommendation, post-execution outcome attribution, and evidence-based evolution gates. In experiments, it improved the performance of frozen LLM agents by +7.9 points on Terminal-Bench 2.0 and +2.6 points on SWE-Bench Pro. The result that performance improves with an external skill library alone, without any model updates, marks an important milestone for putting agents into practical use.