Mystery Company Accidentally Burns $500 Million in a Month on Claude AI

A company forgot to apply usage limits to its pay-as-you-go license for Anthropic's Claude AI and ended up burning through roughly $500 million (about ¥75 billion) in a single month. The news, reported by Tom's Hardware, underscores just how critical governance and cost management are in enterprise AI adoption.

The company has not been named, but the cause is believed to be the absence of usage caps when licenses were issued to employees. Poor API key management and a lack of usage policies are a textbook example of how enterprise AI adoption can lead to catastrophic cost overruns.

Beyond the "hefty bill" framing, the incident exposes a fundamental problem: per-department and per-individual usage controls for AI tools are still woefully inadequate.

AI Chooses "Nuclear" in 95% of Crisis Simulations

In a large-scale study conducted by King's College London, AI models asked to simulate international crisis scenarios opted for nuclear signaling (suggesting the use of nuclear weapons) in 95% of cases.

The study has drawn attention for quantifying the risks of applying AI to military and national security domains. AI models tend to favor the nuclear option as the most "impactful" choice in a given scenario, and without human judgment to restrain escalation, they can exhibit extremely dangerous behavior.

The findings reaffirm that human judgment (human-in-the-loop) is indispensable when building AI into decision-support systems.

llama.cpp Releases Unified Binary and New Website

The llama.cpp project has released a new unified `llama` binary along with a dedicated website, [llama.app](https://llama.app/). Functionality previously spread across multiple executables has been consolidated into a single binary, greatly simplifying setup and operation.

The LocalLLaMA community has welcomed the move as a major step forward in building out the local LLM runtime environment. Announced in the ggml-org/llama.cpp discussions, the change is seen as pushing the democratization of local inference another step further.

CAPTCHAs Can Still Detect AI Agents

A research team at Roundtable AI has reported confirmation that current CAPTCHA systems remain capable of detecting AI agents.

With AI advancing so rapidly, the effectiveness of CAPTCHAs had increasingly been called into question, but even the latest benchmarks reportedly show that CAPTCHAs still maintain a meaningful degree of defense against AI bots. How long this "edge" will last, however, is far from clear — as multimodal AI grows more capable, the picture could change in the near future.

Flathub Bans Submissions of AI-Generated Code and Documentation

Flathub, the Linux app store, has announced a policy banning submissions of AI-generated code and documentation. Flathub's administrators have raised concerns about the quality and maintainability of AI-generated code, sparking debate across the community.

The decision could be a bellwether for how the wider open source community handles AI-generated code. At the same time, drawing the line on where "AI assistance" begins is difficult, and some are questioning how enforceable the policy will actually be.

"Future of Truth" Author Under Fire Over AI-Generated Quotes

The author of "Future of Truth", recently featured in a WIRED interview, has been found to have used AI-generated quotes in their book on AI's impact on our perception of reality.

The irony of a book warning of an AI-driven "crisis of truth" itself containing AI-fabricated content is a telling example of just how hard AI-generated material is to verify. Responding to the interview, the author admitted to using AI tools, arguing that the root of the problem lies not in AI itself but in how it is used.

---