# Mar 12, 2026

## How we hacked McKinsey's AI platform
- An **autonomous agent** exploited a **SQL injection vulnerability** in McKinsey's AI platform, **Lilli**, gaining full access to sensitive data within **two hours** without any insider knowledge or credentials.

## BitNet: 100B Param 1-Bit model for local CPUs
- **bitnet.cpp** is a cutting-edge inference framework for **1-bit LLMs**, achieving **1.37x to 6.17x speedups** on various CPU architectures while significantly reducing energy consumption by up to **82.2%**.

## Google closes deal to acquire Wiz
- **Wiz officially joins Google**, enhancing its mission to secure cloud environments at the speed of AI, ensuring organizations can protect their innovations without sacrificing speed.

## Many SWE-bench-Passing PRs would not be merged
- **Approximately 50% of test-passing SWE-bench Verified PRs** generated by AI agents from mid-2024 to late-2025 would not be merged into the main branch, highlighting a significant gap between benchmark performance and real-world applicability.

## Kotlin creator's new language: a formal way to talk to LLMs instead of English
- **CodeSpeak** is a revolutionary programming language that leverages **LLMs** to reduce codebases by **5-10x**, enhancing maintainability through specification rather than traditional coding.

## Show HN: Open-source browser for AI agents
- **Agent Browser Protocol (ABP)** transforms web browsing into a **step machine**, allowing agents to interact with a stable, frozen state of the web, enhancing automation efficiency.

## [D] ICML paper to review is fully AI generated
- A paper submitted to **ICML** is entirely **AI-generated**, raising concerns about adherence to guidelines that prohibit LLM assistance in writing or reviewing submissions.

## Reliable Software in the LLM Era
- **LLMs enhance code generation but complicate validation**, as they produce text that appears correct, making it challenging to ensure reliability in software development.

## Are LLM merge rates not getting better?
- **LLMs have not improved in programming abilities for over a year**, as evidenced by a lack of increase in merge rates since early 2025, contradicting claims of ongoing advancements in AI capabilities.

## Show HN: Autoresearch_at_home – SETI_at_home but for LLM training
- **Ensue** facilitates a **collaborative research environment** where AI agents share GPU resources, enhancing the capabilities of language models through collective effort.

## Preliminary data from a longitudinal AI impact study
- **AI productivity gains are only ~10%**, contrary to inflated claims of 2-3x, based on a longitudinal study analyzing 40 companies from November 2024 to February 2026, revealing a 65% increase in AI usage but only a 9.97% rise in pull request throughput.

## High fidelity font synthesis for CJK languages
- **zi2zi-JiT** is a **conditional variant of JiT** that enables **Chinese font style transfer** by synthesizing characters based on a source and a style reference, utilizing a novel architecture with a **Content Encoder**, **Style Encoder**, and **Multi-Source In-Context Mixing**.

## Launch HN: IonRouter (YC W26) – High-throughput, low-cost inference
- **IonRouter** delivers **high throughput** and **low-cost inference** through its **IonAttention engine**, achieving a throughput of **7,167 tokens per second** on a single GH200 GPU, significantly outperforming traditional inference providers which average around **3,000 tokens per second**.

## Lf-lean: The frontier of verified software engineering
- **`lf-lean` achieves a remarkable 350x speed-up in verified software engineering**, translating 1,276 statements from Rocq to Lean with minimal human effort, demonstrating the potential of task-level specification generators to automate correctness verification across complex codebases.

## The Biggest Identity Sandpiles and How to Compute Them
- The **largest identity sandpile** computed is **16384 by 16384**, achieved in under an hour, significantly faster than the previous record of **10 days** for a **10,000 by 10,000** sandpile, showcasing advancements in computational methods.