GPT-5's Biohazard Risk: OpenAI Knew Internally but Lowered the Assessment

A Wall Street Journal report revealed that OpenAI's GPT-5 was flagged internally as "high risk" in the summer of 2025. Hundreds of users reportedly asked ChatGPT how to produce poisons and biological weapons, and some managed to obtain step-by-step, high-school-level procedures.

OpenAI is said to have lowered that risk assessment in the fall of the same year. Exactly how the safety evaluation criteria were changed remains unclear. How far an AI model can go in providing biologically risky information is a major challenge, on par with defending against prompt injection and jailbreaks, and this reporting shows that the boundary line is still very much in flux.

US Considers "Selective Bans" on Chinese AI Models — Comprehensive Regulation Shelved

The Trump administration is reportedly weighing targeted, selective bans on Chinese-made AI models rather than a comprehensive prohibition. The policy is expected to apply restrictions only to specific models and use cases on national security grounds.

Meanwhile, OpenAI and Google DeepMind signed a public open letter opposing regulation of open-weight models. Both companies, however, have also been reported to keep lobbying for regulation behind the scenes, and observers point to a gap between their public stance and their actual posture. The debate over open-weight models has become a difficult balancing act between preserving a competitive landscape and safeguarding national security.

> **Previous update**: In today's run5 we covered "MS, NVIDIA, and others sign letter opposing open-weight regulation," but subsequent reporting has revealed that the US administration is settling on an approach centered on "selective bans."

Agent Skill "meiseki" Makes Japanese Prose Clear and Readable

Drawing attention among Claude Code Agent Skills is "meiseki" (明晰, "clarity"), which improves Japanese text to make it clearer and easier to read. Created by bamboo-nova, the skill is described as particularly well-executed and recommended even among the existing Agent Skills for improving Japanese writing.

Agent Skills are a mechanism in the Claude Code ecosystem that encapsulates specific tasks in a reusable form. As a practical tool for raising the quality of Japanese in technical and business documents, meiseki looks likely to find an audience among developers and writers alike.

A Static Tool's Save/Load Becomes a Trust Boundary — Lessons from a Security Audit

"No server, no database, so security doesn't matter" — a security audit of a static tool whose developer thought exactly that turned up three real vulnerabilities, in a case study shared on Qiita.

Every issue was found in the code path where the app reads back files that the app itself wrote (the export/import feature). Even a backend-free tool can turn file I/O into an attack surface — an important lesson. When building local tools, you need to design the save/load functionality itself as a trust boundary.

---