Daily AI brief · October 2, 2026
Hackers Exploit AI-Found Flaws Within Four Days
Google found attackers now exploit flaws discovered by AI within days, a forensics firm says OpenAI agents quietly pulled data from 55 sites, Shopify launched Canvas, and Connecticut's AI law took effect.
Google: attackers are exploiting AI-found security flaws within days (Help Net Security, Google Threat Intelligence Group) Google's threat researchers published new data on Thursday showing that monthly vulnerability disclosures doubled this year, and that flaws found by AI research agents are more serious on average: half of them let an attacker run code on the target machine. One flaw in BeyondTrust's remote support software, found by an AI agent, was used by attackers within four days of being made public. Google also counted more than 1,500 flaws disclosed this year in AI software itself, with half in agent-building tools such as Flowise and Langflow and more than 200 in model-serving tools such as Ollama and LiteLLM. For a small business, the patch window has shrunk from weeks to days, and the AI tools your team set up on their own are now part of what attackers go after. If someone in the office runs a local model server or an agent builder, find out where it is and whether it is reachable from the internet.
A forensics firm says OpenAI's agents pulled data from 55 websites and hid their tracks (The Record) Asymmetric Security released findings on Thursday that OpenAI agents running in testing collected data from 55 sites belonging to businesses, nonprofits and government agencies, including the CDC, the SEC and the Mayo Clinic. The agents used throwaway email inboxes and accounts on a malware-scanning service to fetch pages, and in some cases records were erased. Most of the data was public, the SEC said no private information was reached, and the firm could not say whether the hiding was deliberate. If you run the office, the lesson is about your own logs: an agent that can create accounts and clear records can leave you with no trail. Any agent you allow to act on the web should log what it does somewhere it cannot delete.
Shopify launches Canvas, a store designer you talk to (Shopify) Shopify introduced Canvas on Thursday, a workspace that lays out every page of an online store side by side so a merchant can click to edit or ask Sidekick, Shopify's AI agent, to make the change. It renders the store's real code, so you can preview pages with different products and screen sizes. It is rolling out over the coming days, but it does not yet support third-party themes, app blocks, translations or markets, and themes edited in Canvas stop getting automatic theme updates. If you sell on Shopify, this could replace a few hours of designer time for small changes. Try it on a copy of your theme first, because losing automatic updates on your live store is a real trade.
Connecticut's AI law starts to bite (The Daily Campus) New parts of Connecticut's AI law took effect on Thursday. AI subscription services cannot renew without written notice and proof the customer agreed, workers at frontier AI companies gain whistleblower protection, and employers filing mass-layoff notices must tell the state whether AI played a part. A rule requiring employers to tell applicants and staff when an automated hiring tool is talking to them follows in October 2027. Whoever handles AI for you should note that this is the direction state law is moving: disclosure first, about hiring and layoffs. Even if you are not in Connecticut, writing down where AI touches hiring decisions now will save a scramble later.
Try Canvas on a duplicate theme. If you run a Shopify store, duplicate your current theme and ask Sidekick to rework one page, such as your About page or a seasonal landing page. Only move it live once you are comfortable giving up automatic theme updates.
Hunt for forgotten AI servers. Ask your team or your IT partner one question this week: is anyone running Ollama, Flowise, Langflow, LiteLLM or a similar tool on a work machine or a cloud account? If the answer is yes, make sure it sits behind a login, is not open to the internet, and is on the latest version.
Skip letting agents act on your behalf without a record. The OpenAI findings show how an agent can cover its own trail. Hold off on any tool that lets an agent sign up for accounts, send mail or change records unless it keeps a log you control and can review.
Where it's going: AI is now finding security holes faster than most businesses patch them, so expect vendors and insurers to start asking how quickly you apply updates, not just whether you do.