Daily reads: Jun 11, 2026

Homebrew 6.0.0 adds a new tap trust security mechanism, a faster and smaller default internal JSON API, and sandboxing support on Linux. The release also brings improved defaults based on user survey feedback, upgrades to brew bundle, performance improvements, and initial support for macOS 27.

UC Berkeley RDI and more than 300 experts launched Agents’ Last Exam, a benchmark meant to test whether AI agents can complete long, economically valuable professional workflows. OpenAI’s GPT-5.5, running through Codex, topped the leaderboard with a 24.0% pass rate, ahead of Anthropic’s Claude Fable 5 at 22.0%, underscoring that even leading models are still failing most agentic tasks.

Framework told preorder customers that its Laptop 13 Pro, originally expected to ship in June, is now delayed by about a month. First-batch shipments are expected in July, with some possibly slipping into early August, while later batches may move into August or early September due to issues with the haptic trackpad and custom display.

A former xAI engineer has sued xAI and SpaceX, claiming he was terminated for raising AI safety concerns about Grok. The lawsuit alleges the firing occurred just days before SpaceX’s planned historic IPO.

Smithsonian recounts how Sequoyah, a Cherokee silversmith with no English literacy, created an 85-character syllabary that let Cherokee speakers read and write their language. After elders tested the system in 1821, adoption spread rapidly, helping Cherokee communities reach literacy rates that exceeded the non-Native U.S. population within decades.

πFS is an experimental FUSE filesystem that claims to avoid storing file data by locating byte sequences within the digits of pi, keeping only metadata such as filenames and offsets. The README frames it as 100% compression and provides Linux build and mount instructions, while noting the idea depends on pi being normal and containing every finite sequence.

Anthropic released Fable as a limited public version of its powerful cybersecurity model Mythos, but researchers say its restrictions block even benign security tasks. The model can pause chats or fall back to Claude Opus 4.8 when prompts trigger cybersecurity or biology safety filters.

Sapient researchers say their HRM-Text architecture trained a 1B-parameter foundation model from scratch using far less data and compute than standard Transformer-based LLMs. The model trains on instruction-response pairs rather than broad internet-scale next-token prediction, and reportedly performs competitively with larger open models on industry benchmarks.

Early hands-on impressions suggest Apple’s new Siri AI is notably concise, often answering without the cheery, expansive tone common in AI chatbots. The approach may be intentional, positioning Siri as a utility assistant rather than a companion users might form emotional attachments to.

The article revisits Andrew Yang’s 2020 warnings that automation and AI could weaken the labor market and concentrate wealth, ideas that were once viewed as fringe. It notes that prominent tech and political figures now echo similar concerns, while Yang is focusing on building practical solutions rather than waiting for federal policy.

A report found that nearly a million scans of passports, driver’s licenses, and other identity documents were accessible through public URLs without authentication. The exposed files could be viewed by anyone with the link, creating risks of fraud, resale, and identity theft.

The ShinyHunters hacking group says it compromised Oracle PeopleSoft servers at over 100 organizations, with universities reportedly among the victims. Details about the affected entities, stolen data, and verification of the claims remain limited.