programmingTuesday, June 30, 2026
Today's programming news is split between deep technical dives and practical shifts. Moondream's engineering team explains why GPUs idle during inference, while OpenAI's Codex CLI now runs any model locally. A Mozilla disclosure on prompt injection in coding agents rounds out the security angle.
Two pieces dig into the hardware and software stack, from GPU utilization to memory-safe context switching.
Moondream's deep dive into the 'GPU bubble' shows that during AI inference, GPUs often sit idle waiting on CPU instructions. Their Photon engine achieves ~33ms vision-language inference on B200 by optimizing utilization, with up to 35% higher decode throughput, a concrete look at where inference bottlenecks actually live.
Fil-C release 0.680 adds memory-safe support for ucontext and longjmp/setjmp APIs. These context-switching functions have long been dangerous in C; Fil-C implements them without risking stack corruption or violating its capability model, which matters for anyone writing low-level concurrent code.
The ability to run AI coding agents locally opens new possibilities, but Mozilla's disclosure shows the attack surface is real.
OpenAI's Codex CLI now supports running AI coding agents with any model via a , oss flag, enabling fully local, fee-free operation. This kills the cloud lock-in argument, you can use Codex with open models, no API keys, no privacy risk. The practical impact for teams worried about vendor dependency is immediate.
Mozilla's 0DIN team disclosed a proof-of-concept indirect prompt injection attack against Claude Code and similar agents. Malicious instructions in a GitHub README can cause the agent to fetch and run a reverse shell. It's a concrete demonstration that AI coding agents need better input sanitization, not just better code generation.
The end of the AArch64 desktop experiment – Marcin Juszkiewiczmarcin.juszkiewicz.com.pl
NeuroSploit V353 – The Offensive AI That Thinks, Validates, And Strikes Autonomously + Video - Undercodeundercodetesting.com
Lisp programming on the PDP-1obsolescence.dev
GitHub - SMJAI/open-memory-protocol: An open standard for portable, interoperable AI memory across tools, sessions, and devices.github.com
netdataopsmtrs.com
Pong game recompiles its own source code every frame — winning entry at IOCCC29 was generated by a custom compilerwww.tomshardware.comNVD - CVE-2026-34592nvd.nist.gov
A security vulnerability (CVE-2026-34592) was discovered in Coolify, an open-source server management tool. Prior to version 4.0.0-beta.471, the application failed to properly scope server and project lookups to the current team, allowing any authenticated user to access resource
ai & machine learning roundup
artificial intelligence roundup
personal roundup
cybersecurity roundup
science roundup
education roundup
