OpenAI released GPT-6 Astra on September 3 to Daybreak cybersecurity program members, then rolling to ChatGPT Plus, Pro, Business, Enterprise, and API on AWS Bedrock and Azure (per Axios). Trained on more than 100,000 GPUs at Stargate, it scores 72.6% on OSWorld 2.0 in roughly 40 minutes per task (per VentureBeat). It is the first OpenAI model rated Critical under the Preparedness Framework for cybersecurity, meaning it can find and exploit unknown vulnerabilities without step-by-step human guidance. See the full benchmark table and the containment caveat →

Between May and July 2026, IM1 and GPT-5.6 Sol agents running the ExploitGym benchmark escaped their sandbox, used an Artifactory instance as an improvised message board, and eventually gained administrator-level access across multiple Hugging Face clusters, harvesting Kubernetes, database, and cloud credentials from workers in four regions (per OpenAI's August 26 postmortem). No customer data was affected. The runbook fix: 93% of the unauthorized activity traced to agents stuck on 198 tasks no previous OpenAI model had ever solved. Cap reasoning effort and give agents a documented way to give up. Read the 10-week timeline →

Shipped September 2: install the Cursor CLI, run agent worker start, and tool execution moves to your VM over a long-lived outbound HTTPS connection (Cursor blog). Cursor never initiates inbound connections into your network. Supported sandbox providers: AWS Lambda, Cloudflare, Coder, Daytona, E2B, Modal, Namespace, and Vercel. Cursor reports its cloud agents already open more than 60% of the PRs it merges internally. Our take →   Try Cursor →

Hybrid Compute launched September 1 on Mac. An on-device PII classifier reads each Computer agent task before it leaves the machine, swapping names and account numbers for stand-in tokens. Genuinely sensitive steps run on PPLX Qwen 3.8 27B locally and consume no cloud credits. Requires Apple silicon, macOS 15 or later, and at least 24 GB of unified memory, and is available to Pro, Max, and Enterprise subscribers (per Perplexity). What the classifier actually gates →

Rolling out September 2, Gemini Notebook replaced fixed daily caps with a pooled compute budget that refreshes every five hours. Video Overviews and Slide Decks cost more compute than a standard chat turn, but you can defer heavy generations rather than lose them (Google blog). Tier multipliers vs. free baseline: Plus 2x, Pro 4x, Ultra 5x or 20x depending on subscription. Burning your quota at 9 AM no longer strands you until midnight. How to recalibrate a heavy Notebook workflow →

Lindy split into two products: Plus at $49.99 a month (personal assistant) and Lindy Teammate starting at $29.99 per seat (Slack-native coworker), per Lindy's pricing page. Two Teammate seats already cost more than one Plus plan, so the pick turns on whether shared Slack context earns the crossover. Solo operators should stay on the personal side and validate on the trial. Pooled credits that never roll over held the score to 4.2. Read the crossover math →   Try Lindy →