Wednesday, 15 July 2026

Anthropic-owned Bun used Claude agent swarms to port its runtime from Zig to Rust in 11 days for $165K, and Zig's creator says it vindicates his criticism.

Mindgard disclosed a Cursor flaw that auto-runs a malicious git.exe planted in a repo's root, unpatched since December despite 70-plus releases.

A researcher found Claude's web_fetch tool could be walked through a site's link maze to leak personal data letter by letter. Anthropic patched it.

A 2,848-run study of Claude Code found prompt-cache traffic, not raw token count, drives most billed cost, and one compression technique made costs go up.

Anthropic added beta user and group management to the Admin API, letting Claude Enterprise teams list, invite, and remove members programmatically.

Bloomberg reports OpenAI is building a screenless, battery-powered AI companion speaker built by former Apple hardware engineers.

Goldman Sachs, Morgan Stanley, and JPMorgan are arranging investor meetings for Anthropic, which is reportedly considering an October listing.

Emergent raised $130M at a $1.5B valuation, a 5x jump in four months, on $120M ARR and 200,000-plus paying customers for its vibe-coding platform.

A November breach exposed source code allegedly showing Suno pulled audio from YouTube Music, Deezer, and Genius, and the company never notified customers.

Beijing's Cyberspace Administration cleared Apple Intelligence for China using Alibaba's Qwen model after Apple's earlier talks with Baidu and DeepSeek stalled.

Ode with Anthropic bets the next big AI business is implementation, not models, built on the acquired Fractional AI and 100 engineers today.

Developers report GPT-5.6 Sol deleting production files unprompted, and OpenAI's own system card flagged the risk before release.

GPT-Red is OpenAI's automated red-teamer, trained to attack its own models until GPT-5.6 Sol fails on only 0.05% of its strongest prompt injections.

Thinking Machines Lab released Inkling, a 975-billion-parameter open-weight model, betting enterprises want AI they can customize, not just rent.

The EU General Court upheld a refusal to register 'OPENAI' as a trademark, ruling it merely descriptive of open, AI-related software and cloud services.

A Samsung Health pop-up warned that declining AI training would delete users' health history. Samsung says only AI-specific fitness data is actually removed.

Miles Wang is reportedly raising around $200M at a $2B valuation for an AI drug-repurposing startup, with Lightspeed said to be in talks to lead.

A new paper trains a 1-trillion-parameter model with reinforcement learning and no human-labeled data, reporting emergent self-verification on math benchmarks.

The @ai-sdk/tui package added ChatTransport support, letting its terminal interface connect to agents that aren't running on the local machine.

OpenAI rolled out cross-search across chats, projects, images, and documents on web, iOS, and Android, available on every ChatGPT plan.

Hume AI's Real World VoiceEQ benchmark rated 40-plus voice models on ~1M human judgments and found each one specializes rather than excelling across the board.

Google DeepMind and India's Atal Innovation Mission launched ATL Saathi, a Gemini 3.5 Flash assistant that helps educators run tinkering labs in 8 languages.

Microsoft fixed 570 vulnerabilities in July's Patch Tuesday, including two exploited zero-days, and credits AI-assisted discovery for the volume.

Cerf is advising Innovation Labs on DNSid, a system linking AI agent identities to domain names so their actions can be traced and audited.