Glonce

Edition of

GnuPG patches critical GPGME double free; RCE not ruled out

GnuPG released GPGME 2.3.1 to fix a critical double free that a crafted OpenPGP key with a very long user id can trigger. The advisory says remote code execution can't be ruled out and urges an immediate update. →

OpenRouter off-boards three providers over data policy failures

After security researchers reported that some inference providers had data compromised or were not upholding their data policies, OpenRouter off-boarded three providers and removed 125 endpoints from Zero Data Retention eligibility. →

Zero C# forks bflat to compile C# with no runtime or GC

Zero C#, a fork of bflat, compiles C# into small native programs with no .NET runtime and no garbage collector. The same source targets Linux, macOS, Windows, WebAssembly and DOS on a 386. →

OpenAI publishes GPT-6.2 Astra and Sol prompt templates and API changes

OpenAI released copy-ready prompt templates for GPT-6.2 Astra and Sol covering writing style, initiative and coding, and told developers to review their system prompts. The Responses API now discards earlier reasoning in some cases, and temperature and top_p are gone. →

Anthropic's IPO puts its safety-first pitch under pressure

In a Verge Decoder interview, senior AI reporter Hayden Field argues that Anthropic's Nasdaq listing, with roughly $900 billion in compute bills ahead, will test whether its safety branding survives pressure to grow revenue. →

UN talks on 'killer robots' end with only a deal to discuss options

After a largely fruitless week in Geneva, countries agreed to 'start discussing options' for rules on autonomous weapons. The Red Cross had wanted binding rules drafted within a year; the compromise gives two years to propose regulations. →

ZeroVLN navigates without fine-tuning its VLM brain on navigation data

ZeroVLN is a slow-fast vision-language navigation framework that skips navigation-domain supervised fine-tuning of its VLM brain. The authors say it beats SFT-based hierarchical methods on VLN-CE benchmarks while barely changing the brain's weights. →

GTA 6 hits $3.8bn in first-day retail sales, Take-Two says

Take-Two, parent of Rockstar Games, estimates Grand Theft Auto 6 reached $3.8bn (£2.9bn) in retail sales on its first full day, which the BBC calls the biggest entertainment launch in history. The figures are company estimates through end of day Thursday. →

AutoMaster agent beats vanilla SFT by 1.42-8.99 points with Skill-FT

A preprint proposes Skill-FT, which uses natural-language skills to steer a teacher model's rollouts before fine-tuning, and AutoMaster, an LLM agent that designs those skills. The authors report mean gains of 1.42-8.99 percentage points over vanilla SFT. →

Wired: AI bug flood strains patching as outages hit Microsoft, Apple, NSA

Wired's Kernel Panic newsletter says last week's outages were not an AI hacking spree and have routine explanations. But it argues AI-driven bug discovery is pushing patching to the limit and starting to play a role in real-world failures. →

FlashMorph2 picks which layers keep full attention in hybrid models

A new paper introduces FlashMorph2, an on-policy method for choosing which layers of a pretrained Transformer keep full attention when it is converted to a hybrid. It reports gains over FlashMorph on RULER, LongBench and recall-intensive tasks. →