Anthropic's Initial "Project Glasswing" Report — Unreleased Model Finds Over 10,000 Vulnerabilities

Anthropic has published an initial report on Project Glasswing, a cybersecurity project leveraging its unreleased AI model, Claude Mythos Preview. Working with around 50 partner companies, the project reportedly discovered **more than 10,000 critical vulnerabilities** in the world's important software in just one month after launch.

The project's scale and speed are extraordinary. Findings that could have taken years with conventional security auditing tools and human researchers were produced comprehensively by an AI model in a very short time. While Mythos Preview has not been publicly released, it is believed to deliver exceptionally high performance in code analysis and vulnerability detection.

However, the report also flags a serious problem: **patching of the discovered vulnerabilities is not keeping pace with the speed of discovery**. It exposes an asymmetry — AI's ability to find vulnerabilities is improving rapidly, while the human resources and processes required to fix them cannot keep up. This challenge is likely to become even more pronounced as AI-driven security research ramps up in earnest.

New "Flex" and "Priority" Tiers Added to the Google Gemini API — Flex Available at Half Price

Google has added new service tiers, Flex and Priority, to the Gemini API. The headline: **the Flex tier is available at half the price of the standard tier**.

API costs have long been a major burden for developers, and for applications that consume massive amounts of tokens, monthly bills running into thousands of dollars are not uncommon. The Flex tier appears to slash costs significantly by trading away a bit of response-time immediacy.

The Priority tier is the opposite — an upper tier that processes requests with higher priority. It seems aimed at production environments where real-time responsiveness matters. With these two new tiers, developers can now flexibly choose the cost-performance balance that fits their use case.

The API price war among providers is intensifying, and this move by Google amounts to a clear pricing offensive against competitors such as OpenAI and Anthropic.

Shard — Compressing the KV Cache 10×, Lightening Long-Context Processing for Local LLMs

A technique called Shard that compresses the LLM KV cache by **roughly 10×** is drawing attention in Reddit's LocalLLaMA community.

Shard is a drop-in replacement for HuggingFace's cache mechanism, reportedly achieving about 10× memory reduction at 8K context and about 11× at 32K context on Llama-3.1-8B. Crucially, it shows **no measurable accuracy degradation** on the NIAH (Needle in a Haystack) benchmark or LongBench.

The technical key point is that it takes different compression approaches for Keys (K) and Values (V). It applies PCA + int4 quantization to the K matrix (exploiting the fact that it becomes effectively low-rank once RoPE is removed), and combines Hadamard rotation + vector quantization for the V matrix. Notably, it performs attention computation directly on the compressed K, eliminating the need to restore it to fp16.

For users running large models locally, long-context memory pressure has been one of the biggest barriers. Techniques like Shard could make LLM use on consumer hardware considerably more practical.

Foundation Protocol — A Proposed "Coordination Layer" for a Society of Autonomous Agents

Foundation Protocol (FP), published on Hugging Face Daily Papers, is research proposing the **coordination layer** required for autonomous AI agents to function as societal infrastructure.

Autonomous agents are no longer mere tools — they are transitioning into "societal infrastructure" that browses, makes purchases, deploys software, administers systems, and interacts with other agents. The authors of FP argue that at this scale, the bottleneck shifts away from raw model capability toward **coordination**.

FP provides a graph-based coordination layer that handles agents, tools, resources, humans, and organizations in a unified way. It bakes in economic primitives (metering, receipts, settlement), along with policy, provenance, and auditing as first-class concerns. Rather than replacing existing protocols, it is designed to wrap and bridge them, enabling incremental adoption.

As agentic AI moves toward real-world deployment, the question of how multiple agents coordinate and are governed cannot be avoided. FP deserves attention as a blueprint for the infrastructure that will underpin it.

LLMs Are Overthinking — What Research Quantifying Reasoning Redundancy Revealed

How Much Thinking is Enough?, a paper published on arXiv, is a fascinating study that quantitatively measures "overthinking" in reasoning LLMs.

Reasoning-capable LLMs generate long chains of thought (CoT) when solving hard problems, which carries substantial costs in latency, GPU time, and energy. The research team defined "reasoning redundancy" by how many steps can be deleted from the end of a correct reasoning trace while the model still answers correctly.

Large-scale measurements across four frontier reasoning models and two math benchmarks revealed that **step-level redundancy reaches 61% to 93%**. In six of the eight model-benchmark conditions, only a single step was needed to reach the correct answer.

More importantly, the paper proves that this redundancy is not a flaw of individual models but a **structural property**. As long as length-agnostic outcome rewards are used, no finite stopping time is optimal. Regardless of the RL algorithm, base model, or data distribution, overthinking is not "a bug to be fixed" but "a structural consequence of the current training paradigm."

This finding carries important implications for making reasoning models more efficient, suggesting that the training paradigm itself — not just the models — may need to be reexamined.

EMNLP 2026 Submissions Surge to 11,000 — Up 38% From Last Year

Reddit's MachineLearning community is abuzz over the news that EMNLP 2026 submissions have already reached **11,000**. Last year's EMNLP drew about 8,000 submissions, making this an increase of roughly 38%.

Across AI — not just NLP — top conferences continue to see submissions grow year after year. NeurIPS, ICML, and ICLR show the same trend, reflecting the rapid expansion of the research community. But if growth continues at this pace, challenges such as maintaining review quality and declining acceptance rates could become more acute.

This EMNLP figure is a symbol of how interest in AI research is still expanding at an accelerating pace.

---

**References** (sources for this article):

[Anthropic's "Mythos Preview" Finds Over 10,000 Vulnerabilities in One Month (ITmedia AI+)](https://www.itmedia.co.jp/aiplus/article/2605/24/2000000019/)

[Gemini API's Half-Price "Flex" Tier (@IT)](https://atmarkit.itmedia.co.jp/ait/articles/2605/26/news039.html)

[Shard - getting to 10× KV cache compression (Reddit LocalLLaMA)](https://www.reddit.com/r/LocalLLaMA/comments/1tnvo7r/shard_getting_to_10_kv_cache_compression/)

[Foundation Protocol: A Coordination Layer for Agentic Society (Hugging Face)](https://huggingface.co/papers/2605.23218)

[How Much Thinking is Enough? (arXiv)](https://arxiv.org/abs/2605.23926)

[Already 11,000 submissions for EMNLP? (Reddit ML)](https://www.reddit.com/r/MachineLearning/comments/1tnwdss/already_11_000_submissions_for_emnlp_d/)