« All posts

» Summary

3–Aug 9, 2026

3–Aug 9, 2026
This week

Shai-Hulud npm Worm Infects 1,280+ Packages via Compromised Keyv Maintainer Account

A fast-moving supply-chain attack compromised the npm package Keyv and hundreds of others, exposing developer workstations and CI systems to credential-stealing malware. Attackers used the GitHub account of Keyv maintainer Jared Wray to publish version 6.0.0 via GitHub Actions, bypassing safety verification despite valid provenance on npm. Dubbed a Shai-Hulud variant by Aikido Security, the worm spread to over 1,280 packages; affected teams should remove tainted releases, rotate credentials, and scan for unauthorized access.

In other security news, researchers uncovered a Linux kernel vulnerability in the TC packet editor (`tcf_pedit_act()`) that enables page cache corruption and root access in under a second. The exploit, named “Dirty Pedit,” leverages a u32 integer wraparound to bypass `skb_ensure_writable()` protections, allowing direct writes to page-cache-backed socket buffers.

AI safety took center stage with the introduction of the Minimal Standard for Safeguards v1.0 benchmark. Testing Claude Fable 5, GPT-5.6 Sol, Gemini 3.1 Pro, and Grok 4.5 against 67 jailbreak techniques and 360 high-risk goals revealed a 100x gap in robustness between models. Random search found 63 universal jailbreaks—prompts effective on over 75% of targets—highlighting inconsistent safeguards across frontier systems. Meanwhile, the cost of persistent agent memory emerges as a write-path challenge: each turn triggers LLM calls for extraction and deduplication, with extraction alone consuming 60–75% of memory spend.

Benchmark integrity also came under scrutiny. An audit of the SciCode scientific-coding benchmark uncovered 263 defects across 91% of its problems, falsely penalizing correct, instruction-following solutions and masking true model capability. Similarly, Alibaba’s Qwen 3.8-Max saw its claimed coding performance undercut when independent testing with tighter time budgets (45–60 minutes vs. up to 12 hours) dropped it from top-tier to mid-pack, underscoring how latency allowances can flip ranking tables.

For developers, PHP 8.6 (due November 2026) previewed partial function application, a Polling API for unified I/O multiplexing, property hooks on interfaces, and default values for readonly properties. A coding agent from RapidNative gained attention for using Lifo, an open-source TypeScript Linux engine, to run agent tool calls entirely inside a Vercel Function without VMs or sandboxes, leveraging a simulated filesystem and shell. On iOS, a deep dive into standalone PWAs confirmed that the `interactive-widget=resizes-content` meta tag is ignored by WebKit, forcing developers to use two-finger scroll detection to position chat composers above the keyboard.

Finally, hardware enthusiasts reverse-engineered Nvidia B200’s `tcgen05.mma` tensor-core instruction to bit-level accuracy by probing real silicon. The resulting software model corrected naive IEEE rounding assumptions that failed on 144 of 2,000 test cases, offering the first complete behavioral specification of the instruction beyond Nvidia’s sparse PTX documentation.

» Statistics

Posts
121
Reads
3
Avg. score
7.6

» Most read

  1. Tracking Down a Zsh History Data Loss Bug17.4
  2. Privacy Is No Longer a Requirement, It's an Architectural Constraint17.0
  3. Normalization Problem in Prediction Markets: Understanding Price Differences17.0
  4. DeepSeek-V4-Flash 0731: Full Precision Lossless Performance Test07.3
  5. Pier: Give Each Agent Session Its Own Cloud VM on Your Infrastructure07.5
  6. Model Genome: Fingerprinting Whether an LLM Was Trained from Scratch or Derived07.4
  7. Benchmarking LLMs in File System Design and Implementation07.5
  8. Three.js gets native Gaussian Splatting support08.1
  9. JustAPI: A Rust-Based Python Web Framework07.5
  10. A Tool to Quickly Audit LLM-Generated Python Codebases07.4

» Top scored

  1. Reverse-Engineering Nvidia B200's Tensor Core Down to the Bit09.1
  2. Linux Page Cache Vulnerability via TC Pedit: A New Exploit08.9
  3. Coding Agent Skips VMs and Sandboxes With a TypeScript Linux Engine08.7
  4. AI Jailbreak Benchmark Reveals 100x Safety Gap Between Models08.6
  5. PHP 8.6 Preview: Partial Application, Polling API, and Clamp()08.5
  6. SciCode-Verified: Flawed Benchmark Masked True LLM Coding Skill08.3
  7. iOS standalone PWA: keeping a chat composer above the keyboard08.3
  8. Qwen 3.8-Max and Claude Opus 5 show why benchmark scores don't predict cost08.3
  9. Shai-Hulud npm Worm Infects Over 1,280 Packages08.3
  10. Persistent AI Agent Memory Is a Write-Path Cost Problem08.3

» Sources

Hacker Noon17Hashnode #1015Hashnode #914Hashnode #1512Hashnode #137Hashnode #37The Register6Hashnode #116Artificial Intelligence Reddit5VentureBeat5Hackread5Hashnode #174Programming Languages Reddit3Hashnode #162ÇözümPark2Frontend Development Reddit2Hashnode #191Hugging Face Blog1İşletim Sistemi Reddit1Meta Engineering1Netflix TechBlog1Hashnode #11Database Reddit1Cyber Security Reddit1Webrazzi Yapay Zeka1

» Share