OpenAI Under Investigation by State Attorneys General — Advertising and Health Data Practices in Focus

According to a TechCrunch report, OpenAI is under investigation by attorneys general from multiple states. The specific states involved have not been named, but the investigation reportedly spans a broad range, from OpenAI's advertising policies to its handling of health data.

Details remain limited, and the specifics of the investigation and its outcome are still unknown at this point. That said, multiple states joining forces to scrutinize the data practices of a major AI company is consistent with the regulatory tightening trend seen since 2025. Alongside Europe, where the EU AI Act is being phased in, this suggests that state-level AI regulation in the US is beginning to have real teeth.

US Government Halts Powerful AI "Mythos" — The Impact of Fable 5

Multiple sources report that the US government has moved to halt operations of the powerful AI model "Mythos." According to a report from Pentesty, the measure targets systems known as Fable 5 and Mythos 5, citing cybersecurity concerns.

Commenting on the case, an analysis from grith.ai argues that "the Mythos incident proved that it is no longer possible to guarantee AI safety within the model itself." There is also a growing view that know-your-customer (KYC) measures for AI are insufficient to stop cybercriminals.

The impact has reached actual development teams as well. An article on Zenn describes how a developer who had built their own coding agent running on local LLMs was left scrambling after Fable 5 suddenly became unavailable. The risk of cloud-dependent AI tools suddenly becoming unavailable has become a reality.

AI Benchmarks Are Turning Into "Emissions Tests," Some Argue

An analysis piece from Signal Memo is drawing attention. The article compares the practice of AI companies optimizing their models to score well on benchmark tests to the automotive industry's emissions cheating scandal (Dieselgate).

The argument: benchmark scores keep improving, yet they increasingly diverge from performance in real-world use cases. As confidence in how AI models are evaluated comes under question, calls are growing for independent evaluation bodies and a fundamental redesign of the benchmarks themselves.

The Dangers of AI-Generated Code Made Visible — A Security Scanner Developer's Experience

In an article published on Zenn, the developer of a security scanner reports on the reality of AI-generated code. They scanned code produced by Copilot, Cursor, Claude, and ChatGPT across the board, and looking back, described it as "far worse than I imagined."

Hardcoded API keys and missing input validation were among the issues most frequently found. With more code flowing into production without going through a review process, the author argues that security tools designed specifically for AI-generated code are needed.

In a separate Hacker News post, an audit of 162 PRs created by AI agents found that 27% were cases of "AI fixing its own mistakes." While AI continues to streamline development, questions about quality control are being raised once again.

EU AI Act Article 50 Transparency Obligations Take Effect in August

An article on Zenn summarizes compliance with Article 50 of the EU AI Act (transparency obligations), which takes effect on August 2, 2026, from an engineer's perspective. The provision requires disclosure of AI-generated content, clear labeling of chatbots as AI, and marking of deepfakes, among other things.

The article presents a flowchart for determining in three minutes whether your service is covered, and introduces the minimum necessary compliance measures. With less than two months until enforcement, Japanese companies operating services with EU users may also be required to comply.

---