OpenAI Admits It: AI Hallucinations Are "Mathematically Inevitable"

OpenAI has officially acknowledged that LLM hallucinations (plausible-sounding falsehoods) are not merely an engineering flaw but are mathematically unavoidable. According to the company, the problem is fundamental, rooted in the probabilistic nature of generative models themselves, and a complete solution is considered difficult even in theory.

This statement carries significant implications for companies deploying AI in high-reliability domains such as business operations and healthcare. It effectively sets a theoretical limit on the optimistic view that treats hallucination as merely "a near-term technical challenge."

Kidnapping Scam Using AI Voice Cloning — Bay Area Mother Loses Thousands of Dollars

A mother in California's Bay Area received a call from scammers who used AI to clone her daughter's voice and was defrauded of thousands of dollars. The case has been reported as a real-world example of a "virtual kidnapping scam," in which perpetrators stage a fake kidnapping scenario.

Advances in AI voice synthesis have made it possible to produce highly convincing imitations from just a tiny voice sample, and this type of fraud is surging as a result. Scams that impersonate a family member's voice to fake an emergency show that as the technology becomes democratized, the risk of abuse is expanding alongside it.

Xiaomi Cuts AI Costs by Up to 99% in the Wake of DeepSeek

It has been reported that China's Xiaomi has applied DeepSeek's technical approach to cut its AI operating costs by up to 99%. The move pushes even further along the path DeepSeek demonstrated: achieving performance rivaling large-scale models with a small number of hyperparameters.

The dramatic reduction in AI costs is expected to accelerate AI adoption, particularly in the Chinese market. At the same time, some have raised concerns that cost-cutting could lead to shortcuts in security and quality assurance.

Critical Vulnerability "BadHost" in the Python AI Stack

A vulnerability dubbed "BadHost" has been discovered in Python web frameworks including Starlette, allowing host-based security checks to be bypassed with a single character. It has the potential to affect the entire Python AI stack, posing a serious risk to organizations that use Python as the backend for their AI applications.

Anthropic's Agent Containment Design Draws Attention

An article titled "How we contain Claude across products," published on Anthropic's engineering blog, has been explained on Qiita and is also attracting attention in Japan's engineering community. As agents become more autonomous, the article shares practical insights on controlling the "blast radius," including designing permission scopes for agents and sandboxing them.

---