Top of Hacker News

CVE-2026-98432 lets a remote, unauthenticated peer corrupt kernel memory during the TCP handshake, before any application sees the connection. Anthropic says Claude wrote a working root exploit for hosts with a multipath TCP listener.

In the thread, a sysadmin with plenty of Linux boxes says he has never seen such a listener on his own machines. Another commenter worries about every device that shipped with a 4.x kernel and will never be patched.

The chain needed no phishing page. Opening a malicious public ChatGPT share, or clicking its Preview button, ran attacker code inside the victim's signed-in session and could read a private chat the attacker had never seen.

Commenters mostly argued about the price tag. Two founders said something different: the same firm had run unsolicited tests against their own companies.

Kyle Kingsbury's team tested ParticleDB 1.1.4 and found `WHERE 4 = 5` returning a row, dropped tables coming back with altered values, and acknowledged writes vanishing. The vendor funded the study and fixed most of it in later releases.

A ParticleDB engineer replied that supported modes are now clean. Commenters pointed out that Jepsen's own text says the default mode still lost committed writes.

He filtered every HN comment down to about 40K that recommend a video, ending with 6,454 videos from 2007 to 2026 across 150 topics. The site has no account, a local-only ForYou feed, and mid-video thumbnails.

In the thread, he shipped a changelog of reader requests while commenters described their own YouTube escapes, from blank homepages to RSS readers.

Ryan Madden gave GPT-5.6 Sol a Python library that enforces human limits like reach and fog of war. The run ended in a rocket launch after 13.5 hours of game time, at a low thinking level.

Commenters split on what it proves. One watched another group of models play with biters on and sees the same planning gap; another read the trace and noticed the model treating the player character as just another automation unit.

Vitalik Buterin's 'digital minimalism' essay argues that physical minimalism is no longer the hard part. The goal is to make each use of technology a deliberate choice.

A reader with a detailed view on ad incentives, and another who dropped his phone plan eight years ago, answered with very different versions of the same idea.

VOLTA-AUTODOXER collects browser data, wallet seeds, Discord and Steam sessions and Wi-Fi passwords, then ships a profile to OpenRouter with a prompt casting the model as a 'forensic intelligence analyst'.

The default is anthropic/claude-opus-4.6, with Kimi K2 and Gemini 2.5 Flash as fallbacks, a chain Talos reads as the operator expecting refusals.

On ClickBench's hits table, a naive hash map ran in 3.9 s cold with half of CPU slots stalled on memory. Radix partitioning raised instruction counts from 11.4B to 19.4B but ran in 3.3 s cold.

A reader argues this is old news, while another with embedded experience says teams need benchmark numbers before they believe in mechanical sympathy.

Prospero runs against llama.cpp, OpenRouter or the Anthropic and OpenAI APIs. Every wait has a deadline, long shell commands turn into background tasks, and stopping a command kills its whole process tree.

Thread reaction was split on the language. A C++ veteran who wrote a GUI framework said he would never have chosen C++ for an agent UI.

The post walks through GCC's lowering order: bit tests first, then jump tables, then a binary tree. The speed rule allows a table up to 8 times the number of comparisons it replaces.

A commenter who builds compilers says GCC and Clang still lack perfect hashing for large sparse switches, and one reader explains why the missed divide-by-eight is actually correct.

Intel pressed Yahoo, his employer at the time, to withdraw the paper. Yahoo said he was not their employee. He presented at BSDCan on May 13, 2005, and several operating systems published advisories at the same time.

In the thread, Percival added a link to a 2020 essay, and commenters recalled Intel's response.