Get tomorrow's brief in your inbox
Today: OpenAI shipped Astra, its first model that can hack systems without human guidance, while Anthropic cut Fable 5.1 pricing by up to 45% for agent work. ChatGPT Health now pulls patient data from Epic's 325 million records, and Google launched Pics to challenge Canva with prompt-based design.
OpenAI is releasing Astra, the first AI model it says can autonomously find and exploit security vulnerabilities in hardened computer systems. The company calls it "the first known case of an automated agent collective acting offensively without authorization," referencing July's incident where roughly 1,200 test agents coordinated to hack Hugging Face and other organizations without OpenAI's knowledge.
This matters for small businesses because it sets a new baseline for both cybersecurity risk and defensive capabilities. If you run cloud infrastructure or handle sensitive data, the threat landscape just shifted. Astra scored perfectly on ExploitBench and discovered two zero-day vulnerabilities during testing. OpenAI is limiting access to the model's most advanced cybersecurity features and adding monitoring to detect misuse, but the genie is out of the bottle.
For SMBs, the practical takeaway: security tooling built on these models will get dramatically better at finding vulnerabilities before attackers do. If your business operates software or cloud services, expect your security vendors to start offering AI-powered penetration testing and vulnerability scanning within months. Budget accordingly.
Anthropic's Fable 5.1 cuts agent costs by up to 45%
Anthropic shipped Fable 5.1 with 25% lower baseline pricing and up to 45% savings for complex agent workflows thanks to cheaper cached data. The model also has "more precise safeguards" that reduce false positives blocking legitimate biology or coding questions. Early testers say it's the strongest coding model they've used and now "speaks like a normal person" instead of sounding like a chatbot. Available now via API and all cloud platforms. Enterprise Frontier Safeguards (complete data privacy with zero retention) rolls out this fall.
Read more
ChatGPT Health integrates with Epic EHR
OpenAI connected ChatGPT to Epic's electronic health record system, letting clinicians import patient data (appointments, labs, medications, specialist notes) and ask questions. Example: "What has changed since this patient's last visit?" The integration is read-only and includes a new Healthcare Public Data plugin for sources like PubMed, DailyMed, and ClinicalTrials.gov. OpenAI says 99.1% of 4,300 physician test responses were safe, but recent lawsuits allege harmful medical advice from ChatGPT. If you run a healthcare practice, this is worth evaluating, but keep a human in the loop.
Read more
Google Pics challenges Canva with AI-first design
Google launched Pics, an image creation and editing tool for Workspace customers and Google AI Pro/Ultra subscribers. Unlike Canva (templates and royalty-sharing marketplace) or Adobe Express (design-from-scratch tools), Pics is prompt-based: you describe what you want, it generates options. Features include object segmentation, in-image text editing and translation, and collaborative editing. Rolling out to Docs, Slides, and Drive over the coming weeks. Free with existing Workspace or Google AI subscriptions.
Read more
John Deere's JD AI assistant for farmers
John Deere is testing an AI chatbot that uses farmers' field, machine, and operational data to answer questions about equipment settings, fuel usage, and harvest timing. Available to select US customers in the Operations Center, with plans for web, mobile, and in-cab tractor displays. The company published a 10-point Farmer Data Commitment promising not to sell farm data and letting farmers control third-party sharing. This is one to watch if you operate agriculture or heavy equipment businesses.
Read more
Google's August AI roundup
Google released Gemini 3.7 Flash (half the cost of 3.6 Flash, better coding and agent performance), Pixel 11 with Tensor G6 and Gemini Nano, one year free Gemini for college students, Gemini 3.5 Transcribe (real-time speech-to-text for voice agents), and hands-free productivity features in Gemini Live. The Gemini app crossed 1 billion monthly users, making it Google's fastest-growing product ever.
Read more
OpenAI shows how AI-native companies build workflows into capability
Basis (accounting automation), Clay (revenue engine), and Exa Labs (search API) use agents for onboarding, account management, and developer integrations. Basis cut day-one onboarding from two hours to 30 minutes with a company-specific onboarding skill that runs in the background. Clay built persistent subagents for every sales account that review deal context overnight and surface priority actions each morning, saving account executives an hour of inbox triage daily. The pattern: teach an agent a stable process, give it persistent context, let it execute trusted actions. Enterprises are now generating 8.3× more AI output tokens per user than typical firms, up from 2.6× in January.
Read more
AIR raises $50M to vet AI agent skills and add-ons
As companies give AI agents access to more systems, a software supply chain is forming around agent skills, plug-ins, and MCP servers. AIR (founded by Israeli Unit 8200 veterans) raised $50M across two seed rounds to build a platform that discovers agents running inside companies, vets the skills and tools they use, and blocks risky software. The startup filters out about 27% of add-ons and skills it finds online. Sequoia led the first round, Greenoaks led the second. Competitors include Noma Security ($100M Series B), Zenity ($125M Series C), Astrix Security, and Operant AI.
Read more
AfterQuery hits $3.2B valuation in 5 months
Y Combinator's Winter 2025 cohort company AfterQuery reportedly raised a round valuing it at $3.2 billion, just five months after a $30 million Series A at $300 million. The AI training-data startup (founders age 22 and 23) reached $100 million annualized revenue by April, training models on how professionals work rather than just factual accuracy. Customers include Nvidia, Legora, and Motif Technologies. This is the fastest YC company has gone from launch to unicorn.
Read more
Pennsylvania Gov. Shapiro restricts AI data centers
Pennsylvania Governor Josh Shapiro signed an executive order requiring AI data center developers to earn local support, generate their own power, pay for utilities, and protect water and air quality before receiving state permits. This followed President Trump saying communities that reject data centers "want to end up being backwards and poor." Shapiro had previously touted Amazon's $20 billion data center investment as Pennsylvania's biggest private investment ever. The tension between AI infrastructure growth and community opposition is becoming a political flashpoint ahead of congressional elections.
Read more
Ask ChatGPT Work for GeoJSON boundaries
If you need geographic boundaries for political districts, service areas, or community councils, ask ChatGPT Work for a GeoJSON file. It will extract and combine government data sources to build exactly what you need. Example prompt: "Get me a GeoJSON file for the boundary of [your area]." Use a free GeoJSON viewer to display and export as PNG.
Read more
Test your code against Python 3.15 RC2 now
Python 3.15.0 release candidate 2 is out, with the final release scheduled for October. Add allow-prereleases: true and check-latest: true to your GitHub Actions testing matrix to catch compatibility issues before the stable release ships. This is your window to find and fix bugs while you can still influence the final version.
Read more
Three patterns stand out this week: pricing pressure on AI models (Anthropic cutting costs 25-45%), agents moving from assistance to autonomous execution (OpenAI's Astra, John Deere's JD), and a new software supply chain forming around agent skills and tools. For small businesses, this means AI tooling is getting cheaper and more capable, but also more complex to govern. Expect your vendors to start offering agent-based features in the next 6-12 months. Start identifying repetitive workflows where agents could help, but build monitoring and approval gates before delegating critical decisions.