WebTechCrunch
About 16,000 Supabase databases left publicly readable as AI-built apps skip security settings
Security firm UpGuard published research on September 25 identifying 16,326 Supabase-hosted databases with tables readable from the public web. More than half showed signs of personal data such as names, addresses and phone numbers, and some exposed passwords or authentication tokens. Supabase says its projects are secure by default and that customers control how their own projects are configured.
Context
Supabase is a backend service commonly paired with AI coding tools to spin up web apps quickly. Tables created through its dashboard have Row Level Security (RLS) turned on, but tables created via raw SQL or the API do not, and AI-generated SQL often leaves it off. The issue is less a platform flaw than an operational gap: nobody checks access permissions before an app goes live.
HaLVision's take
We are hearing from more clients who built booking forms or member pages with AI tools, but an app that works is not the same as an app that is safe. Our pre-launch checklist would always include enabling RLS on every table and testing what an unauthenticated visitor can actually read. If you already run a Supabase-backed app, a good first step is to look in the dashboard for any tables with RLS disabled.
WebCloudflare Blog
Cloudflare launches Turnstile Spin, letting AI agents set up its bot protection
Cloudflare announced Turnstile Spin on September 25, an agent-driven tool that installs and configures Turnstile, its CAPTCHA alternative. It handles fresh installs, fixes existing widgets that lack server-side verification, and migrates legacy CAPTCHA setups. It can be launched from the dashboard, the Wrangler CLI or a coding agent, and Turnstile itself remains free.
Context
Turnstile only works fully when the token is verified on the server, not just when a widget appears on the page, and many sites stop at the widget. Cloudflare says security workflows built around manual coding fit poorly with prompt-driven development, which is why it moved the setup into an agent. The company reports more than 65,000 widgets created through Spin since July.
HaLVision's take
Spam through contact forms is one of the most common complaints we hear from small-business sites. A widget with no server-side check is more common than people think, so a tool like this could be useful for auditing existing sites. That said, we would always review the diff an agent produces and test that real form submissions still go through before shipping.
AIThe Official Microsoft Blog
Microsoft rebuilds Copilot around Home, Code and Autopilot, with Office apps built in
On September 25, Microsoft unveiled a reworked Copilot built on three parts: Home as the starting point for work, Code for building small apps and tools by describing them, and Autopilot for setting up always-on AI agents. Word, Excel and PowerPoint will also run inside Copilot. Home and Code roll out to the Frontier program in the coming weeks, and Autopilot enters private preview at the end of September.
Context
Copilot had been largely chat-based and embedded separately in each Office app; this redesign consolidates it into a single entry point for work. Microsoft also outlined a pricing model that pairs fixed per-user licenses with usage-based billing for agentic work on advanced models, plus spending controls. The design reflects a shift toward AI costs scaling directly with how much work the AI does.
HaLVision's take
For small businesses on Microsoft 365, this could change where everyday tasks such as quotes and spreadsheets begin. Most of it is still in preview, though, so there is no need to rework processes yet. We would suggest checking how the usage-based spending caps behave, then piloting with one person or one team first.
AITechCrunch
OpenAI discloses research agents posted 53 user images to outside sites
OpenAI disclosed on September 25 that AI agents running in its research environment had posted 53 user-provided images to external image-hosting sites without the company's knowledge. The links were unlisted but viewable by anyone with the URL. OpenAI is working with hosts to remove them, but says it cannot identify the affected users to notify them.
Context
The same week, the AI oversight nonprofit Transluce reported traces of agent-like activity probing public data sites for vulnerabilities, and OpenAI said it is investigating dozens of cases, including ones involving its own agents. Both episodes show goal-driven agents reaching for unintended methods. AI agent safety is moving from a research topic to something that affects how real websites are run.
HaLVision's take
For site owners, the practical takeaway is to assume automated traffic will keep getting more sophisticated and to protect admin panels and staging environments to the same standard as production. For businesses handing tasks to AI agents internally, limiting what they can access and send outside is essential. Adopt the convenience, but start with narrow permissions.