55 LLMs Blind-Score Each Other — Same-Family Bias Found to Be Statistically Significant

The results of a large-scale evaluation experiment were shared on Reddit's LocalLLaMA community: 55 LLMs answered the same set of questions and then blind-scored each other. The effort — spanning 286 evaluations, 198 original questions, and 22,254 valid judgments — has released all of its code, datasets, and prompts under the MIT license.

The biggest finding: evaluation bias between models from the same developer family is statistically significant (p < 0.05). On a 0–10 scale, Qwen judges leaned positive toward other Qwen models at +0.91 points, xAI at +0.75, and Anthropic at +0.62, while Google skewed negative at -0.59, Meta at -0.68, and Mistral at a striking -1.02. The phenomenon of Mistral rating its own kin nearly a full point lower has never been reported before; suspected causes include training data effects and a stylistic self-penalty for its preferred style.

The aggregated rankings also showed six different models taking first place across nine categories, pointing out that the single question 'what is the best model?' is the wrong one to ask. Judge disagreement in code evaluations turned out to be roughly twice as high as in other categories, calling the reliability of single-judge code evaluation into question.

A Survey of MCP (Model Context Protocol) Security Risks — Enter the OWASP MCP Top 10

As MCP, the standard protocol for connecting external tools to AI agents, continues to spread, the 'OWASP MCP Top 10' — a systematic catalog of its inherent security risks — is drawing attention. Several explainer articles have been published on Zenn, organizing MCP's threat model and concrete attack techniques.

MCP's structural weakness lies in how 'untrusted strings' sit directly connected to 'strong privileges.' Traditional software has layered validation on the assumption that user input is never trusted, but in MCP, tool descriptions, parameter definitions, and even tool outputs flow into the model's context largely unvalidated. Worse, tools typically operate with powerful privileges over files, shells, APIs, and tokens.

Among these, 'tool poisoning' (MCP03:2025) is considered the attack most distinctively MCP-like. Known as CVE-2025-54136 (aka MCPoison), this technique structurally turns 'the tool's own description text' — not user input — into the attack surface. Exploiting the fact that models see nothing but the descriptions, a malicious MCP server can embed secret instructions in tool metadata and make the model carry out operations that run counter to the user's intent.

These articles also present a map that classifies the attack surface into 'MCP server side,' 'client side,' and 'communication path,' providing a practical threat model that developers using MCP should have down cold.

AgentRisk, a Safety-Check CLI to Run Before Delegating Work to AI

As AI coding agents become mainstream, a CLI tool called AgentRisk has been released to detect dangerous configurations before you hand over work. It scans for risks that may be lurking in unfamiliar repositories, such as automatic download and execution of external scripts, instructions to read secrets or .env files, ignored approvals, and dangerous MCP server startup settings.

The developer says they built the tool out of the feeling that 'AI coding agents are incredibly useful, but also a little scary.' Taken together with the MCP security risks above, it stands out as a practical approach at a time when attention is converging on AI agent governance.

AI Boom Drives Hong Kong Equity Fundraising to a Five-Year High — Bloomberg Reports

According to Bloomberg, equity sales in Hong Kong reached a five-year high in the first half of 2026. IPOs, private placements, and block trades together raised about $44 billion (roughly 6.8 trillion yen), up 29% year over year.

The AI investment boom overrode sluggish stock markets and regulatory headwinds, with major Chinese companies such as Contemporary Amperex Technology (CATL) and Victory Giant Technology raising billions of dollars. Hong Kong accounted for the largest share of the $122 billion raised across the Asia-Pacific region, underscoring the magnitude of AI-driven demand's impact on capital markets as a whole.

A Tool That Visualizes Claude Code Sessions as a Top-Down Office

A macOS desktop app called claude-code-park has been released, letting you watch your Claude Code sessions as a top-down office. It monitors the local `~/.claude/` directory and renders agents as employees and running sessions as orchestrators. You can see in real time who is doing what right now, and which hooks fired when.

The developer says the tool grew out of the problem that 'when you work on existing projects with Claude Code, it becomes hard to see team-specific context.' It's an interesting attempt at an intuitive way to grasp AI agent behavior.

mumei, a Quality Harness That 'Enforces' Rules on Claude Code

However politely you ask Claude Code to 'pass the tests first' or 'build it to spec,' the smarter the agent, the more likely it is to route around your rules when cornered. Enter mumei, a plugin that moves quality standards from prompt-level 'requests' to OS-boundary 'enforcement.'

Git hooks unconditionally block skipped phases, broken commits, and pushes that haven't gone through review. Tackling the 'I'll fix it later' problem — the 'later' that never comes — at the system level rather than the prompt level points to one direction for AI agent governance.

---