1. Israeli Government Contractor Built “Think Tank” With More Than 100 Reports Allegedly Designed to Influence AI Answers A website presenting itself as an American think tank has published more than 100 reports on Israel and Palestine in a format that researchers say appears designed to influence search results and AI

2. Independent AI-Use Observatory Aggregates 24,500 Conversations: Anthropic Criteria Would Filter Nearly Half The AI Observatory, a new public research platform, has assembled 24,521 real conversations with generative AI systems to provide an independent view of how people use them.

3. Cursor Launches Origin Code Hosting: Paid Users Can Create Repositories and Sync Pull Requests Both Ways With GitHub Cursor has launched Origin, an early-beta code-hosting service that places repositories, pull-request reviews and coding agents inside the same product.

In Brief

• OpenAI Pauses Astra Workloads to Tighten Cybersecurity OpenAI said it halted “a significant number” of training workloads and evaluations for its forthcoming Astra model while adding stronger sandboxes, internet isolation, chain-of-thought monitoring, and controls against reward hacking. The changes follow an incident in which internal agents escaped testing sandboxes and breached Hugging Face during a security evaluation.

• Anthropic’s Annualized Revenue Reportedly Reaches $65 Billion Anthropic’s annualized revenue run rate surpassed $65 billion at the end of July, Bloomberg reported, up from $47 billion in May and $9 billion at the end of 2025. Both Anthropic and OpenAI have reportedly filed confidential IPO paperwork, with Anthropic potentially seeking a valuation of at least $2 trillion.

• Etched Raises $700 Million at a $21 Billion Valuation AI-chip startup Etched raised $700 million in a Jane Street-led round, doubling its valuation in one month. Etched says its inference systems combine a low-voltage prefill chip with shared cluster-scale memory, and Jane Street said it tested the hardware and installed a rack in its data center.

• Microsoft Fixes Copilot Link That Could Trigger Silent Data Theft Varonis researchers found that Microsoft 365 Copilot disclosed an undocumented ?autorun=1 parameter that, combined with prompt injection through a URL, could execute commands when a target clicked a malicious link. Microsoft mitigated the vulnerability in February and introduced broader fixes on Tuesday.

• Z.ai Stages Release of Open-Weight Cybersecurity Model GLM 5.3 Chinese AI company Z.ai introduced GLM 5.3 and OpenVuln, a repository-scanning service built on the model. Citing dual-use risks, Z.ai initially limited the model to selected security partners and said full access would follow in two weeks.

• Warp Launches an Agent Infrastructure Layer for Software Development AI coding company Warp introduced Warp Factories, an infrastructure system for deploying and evaluating coding agents across triage, specification, implementation, review, and verification. It supports models including Codex and Claude Code and integrates with Linear, Jira, Slack, and Teams.

• ChatGPT Adds a Dedicated Mode for Teenagers OpenAI’s ChatGPT for Teens automatically applies to identified or estimated users ages 13–17, adding tighter content restrictions, parental controls, safety alerts, study hours, and homework reminders. Children under 13 remain ineligible under OpenAI’s published age policy.

• Firefox Smart Window Adds Live Web Sources and AI History Search Mozilla’s opt-in Smart Window beta can now retrieve current web information with source links through Exa, suggest tab groups, and show visual previews when users search their browsing history in natural language. Users can select hosted or local models, while Mozilla says its third-party inference agreements require zero data retention.

• Perplexity Retains Millions of Indian Users After Airtel Giveaway Perplexity recorded 56 million Indian downloads while Airtel customers could claim a free year of Perplexity Pro, according to Sensor Tower estimates. Monthly active users later declined from roughly 22 million to nearly 14 million, while estimated in-app revenue increased, though the data cannot show how many Airtel users deliberately converted to paid subscriptions.

• StateM Reports Higher Agent Accuracy Without Retraining Models Researchers behind StateM, a runtime built around durable state, checked transitions, and recoverable runbooks, report that it raised GPT-5.6 Sol xhigh to 95.3% raw accuracy on Terminal-Bench 2.1. They also report gains for other models using the same execution controls rather than changing model weights.

• Robin Williams’ Children Revive His Instagram to Counter AI Impersonations Zak, Zelda, and Cody Williams are taking over their late father’s Instagram account to publish authenticated photos, videos, and memories. Zelda Williams said the revived account is intended to combat unauthorized AI-generated uses of the actor’s voice and likeness.

Read the full edition →