Over 430,000 AI API Keys Exposed in Public GitHub Code

A project called AI Leak Watch reported detecting 435,608 potential AI API key matches in public repositories on GitHub. The findings shed light on the reality that API keys from OpenAI, Anthropic, Google, and others have been widely exposed through mistaken commits and intentional disclosures.

API key leaks have long been a concern for traditional cloud services, but with AI API keys, abuse translates directly into massive billing damages, and secondary risks such as unauthorized access to training data and user conversations are also a worry. Developers are once again urged to add `.env` files to `.gitignore` and adopt secret scanning tools.

Most Americans Distrust Both AI and the People Managing It

According to The Verge, citing survey data from the Pew Research Center and Gallup, most Americans lack trust not only in AI technology itself but also in the people and organizations managing it.

As AI becomes more deeply embedded in society, this lack of trust could influence how regulation takes shape and how much transparency is demanded of companies. The gap between technological progress and social acceptance looks set to remain a fundamental challenge the AI industry must keep confronting.

llama.cpp b9200 Speeds Up MTP Inference

The new llama.cpp release b9200 improves inference performance for Multi-Token Prediction (MTP). It optimizes a step that previously copied logits for every token in the batch during prompt processing, and leverages the fact that MTP requires only pre-normalization (pre-norm) to drastically cut memory traffic. As a result, prefill performance has improved.

The community has also published benchmarks pairing b9200 with Unsloth's Qwen 3.6 27B MTP GGUF builds. There are also reports that Unsloth's MTP flag, previously recommended, can actually become a bottleneck in strict multi-turn agent workflows—a renewed reminder of how important it is to tune settings to your use case.

Solo Project "TIME" Accepted at ACL 2026

An individual developer reported on Reddit's LocalLLaMA community that their project has been accepted at ACL 2026. "TIME" is a method for training Qwen3 models to produce short, context-dependent reasoning. Whereas conventional reasoning models emit a lengthy thinking block up front, TIME triggers a brief burst of reasoning only when the situation calls for it.

What began as a personal tool built for use in the developer's own Open-WebUI setup somehow ended up as a paper at a top-tier conference. Its distinguishing feature is modeling changes in latent context—silence, gaps, stale assumptions, and so on—using the concept of time. Rather than "making thinking shorter," it takes a "re-think as context demands" approach, which makes it worth paying attention to.

Comparing LLM Performance Across Programming Languages

GertLabs has published the results of a study comparing LLM performance across programming languages, focusing on how the language composition of a benchmark affects results when evaluating model performance.

How the syntactic complexity and paradigm differences of each language affect LLM code generation is an important consideration for practical model selection and evaluation design. In particular, the degree of performance degradation on less common languages and the differences between scripting and compiled languages offer some intriguing insights.

M365 Copilot's Comeback from "Clunker" Status

An ITmedia roundtable article highlighted how the reputation of Microsoft 365 Copilot—once derided as a "clunker" at launch—has improved dramatically in recent years. Three Microsoft MVPs discussed its evolution, pointing to the introduction of GPT-5 as the turning point.

The key to getting value from it, they said, is "moving beyond prompt craftsmanship": rather than laboring over elaborate prompts, the recommended approach is to weave it naturally into everyday work. As enterprise AI matures in practice, we seem to be at the point where user expectations and tool maturity are finally crossing paths.