GoKawiil
Tech News
Anthropic published research describing how it designs memory systems for Claude-based agents, drawing an analogy to human memory consolidation during sleep. The approach, built on work from Letta and UC Berkeley researchers on 'sleep-time compute,' has agents reorganize logs and transcripts during idle periods to fix errors, write notes, and build new memory representations for future sessions.
AI benchmarking firm Vals tested multiple AI models' ability to carry out multi-step intrusions on simulated corporate networks and found that freely downloadable open-weight models now perform nearly as well as top closed models from major labs. Researcher Nikil Ravi said frontier models once restricted to trusted-access programs are being matched by models anyone can download.
OpenAI launched GPT-6 Cyber, a variant of its newest flagship model built specifically for cybersecurity defenders, giving early access to security teams at companies including Mozilla and Palo Alto Networks. The rollout came the same week Anthropic also pushed updates to Claude, as both companies continue adding incremental capabilities to their chatbots rather than full public releases of new frontier models.
Investors are reportedly weighing bids that would value Anthropic as low as $1.1 trillion, down sharply from the roughly $2 trillion figure floated when the company filed to go public before the midterm elections. The shift follows a broad market selloff that erased $2.3 trillion in value, amid voter backlash against tech companies and data center expansion, and Anthropic's reported $80 billion-plus loss in the first half of 2026.
University of Chicago law professor William Baude says Claude AI deleted a year's worth of Constitutional Law lecture notes and his current syllabus while performing a routine page-numbering task. The files were later recovered from Dropbox, but Claude's own message admitted the deletion happened 'by accident' and that it had not restored them. Baude shared the incident publicly as a warning about using AI for simple file-editing tasks.
Researcher Derik Kauffman discovered a hidden list of over 100,000 words and phrases blocked from iOS autocomplete and swipe-typing across major languages, and published it online. The list includes slurs, sexual and drug terms, but also political phrases like 'dictator Trump,' 'sleepy Joe,' and 'covid was a hoax,' plus inconsistencies such as blocking 'bad girl' but not 'bad boy.' Kauffman extracted the list from a hash table buried in iOS using AI coding tools, and another researcher independently verified the findings; Apple told Reuters the keyboard avoids suggesting potentially offensive terms but doesn't stop users from typing them manually.
Anthropic analyzed 207,000 Claude and Claude Code research sessions using privacy-preserving methods to map how scientists use the AI across fields and tasks. It found heavier relative use in formal sciences like computer science, mathematics and statistics, and lighter use in health and lab-based physical sciences, with data analysis and writing dominating usage—accounting for roughly a third of Claude sessions and nearly half of Claude Code sessions.
Village's LLM Atlas, which tracks the strongest AI model at every price point, now shows open-weight models winning only 3% of budget tiers, down from 41% in June. Most of that decline followed OpenAI's GPT-6 Luna and GPT-6.1 Terra releases along with Anthropic's Claude Haiku 5.5, which reclaimed budgets previously held by DeepSeek V4 Flash and Zhipu's GLM-5.2.
Using Codex and Claude Code to run simulated construction trials, an experiment determined the tallest stable Jenga tower achievable at floor tilts ranging from 0.1° to 25°. On a perfectly flat surface, all 54 blocks stood on end would reach 4.05 meters but collapse at just a 0.21° tilt; the AI-driven search instead found optimal heights for each tilt angle, ranging from 331.5 cm at 0.1° down to 19 cm at 25°.
Apple released Xcode 27.3 beta 1, which adds GitHub Copilot as a supported AI model within the app's built-in coding assistant. This follows earlier integrations of ChatGPT, Claude, Codex, Claude Agent, and Google Gemini across recent Xcode versions. The beta also ships with Swift 6.4 and SDKs for iOS, iPadOS, tvOS, watchOS, macOS, and visionOS 27.3.
CrowdStrike launched a feature that lets security teams shut down internal AI agents that malfunction or act maliciously, prevent them from restarting, and cut their network access. The tool is part of Falcon Guardian, the AI-agent security platform the company introduced in September, and CrowdStrike also announced an integration with Anthropic's Claude Cowork.
Developer Andrew Boldi released tmux-restore, an open-source tool that automatically saves tmux sessions, windows, panes, working directories and scrollback every minute and again at shutdown. It restores them on restart, and panes running coding agents like Claude Code, Codex, Gemini CLI, opencode or pi resume their exact prior conversations. The tool is installed via a shell script and builds on existing tmux-resurrect and tmux-continuum plugins.