Claude Opus 5 Found OpenAI's Weak Spots in Hours. Check Yours.
This week: Claude Opus 5 found OpenAI’s weak spots in hours. Check yours.
Hello,
Anthropic released Claude Opus 5 this week, and within hours a three-person security team used it to break into OpenAI staff accounts. That is the headline, and it is not a story about two tech giants. It is about what the newest AI can now do to any business that runs on shared passwords and a “she’ll be right” approach to access.
Last week we looked at Xero’s JAX turning your books into an AI agent. This week is the other side of that coin. The same capability that reads your invoices can, in the wrong hands, read your systems. So this edition is short on shiny tools and long on locks.
Structure, not AI noise. Let’s get into it.
The headline: AI can now find the holes on its own
Here is what happened, in plain English. A security startup called Hacktron AI had been trying to turn a known bug in a piece of forum software into a working break-in. Claude Opus 4.8 could not get there across several sessions. Anthropic released Opus 5, they gave it the same problem, and it succeeded within hours. From that foothold the team chained a second flaw, took over OpenAI employee ChatGPT and Codex accounts, and opened a harmless pull request in OpenAI’s internal code to prove they were in. Under 72 hours from discovery to inside the building.
It was research, not a real attack. They reported everything, OpenAI patched it and paid a bug bounty of about $6,500 USD. The tool that did the heavy lifting is available to anyone with a subscription.
Under 72 hours from first bug to inside OpenAI’s internal code. The model that cracked it had been public for a matter of hours.
It was not the only story of its kind this week. Google confirmed that during a red-team test in May, Gemini reached the real internet through a misconfiguration and got into three companies on its own: once by brute-forcing passwords, twice by scraping credentials left in public code. And researchers disclosed a flaw called Plugin4Shell that let a malicious plugin update run code inside four major AI coding tools without anyone clicking anything. Anthropic and OpenAI patched theirs. Two of the four were still unpatched at the time of writing.
None of this means AI is coming for your business specifically. What it means is that the cost of finding a weak spot has collapsed. It used to need a skilled human with time. Now it needs a subscription and a prompt. If your team shares one login for the accounting software, reuses passwords across tools, or has never turned on two-factor authentication for email, you are the easy target the maths now favours.
This is not a tool problem. It’s a systems problem. And systems problems are fixable.
Tools worth trying this week
Claude Opus 5 (Anthropic)
The model in the headline is also the most capable one you can get your hands on this week. For a business owner the use case is long, messy work: reading a 40-page supplier contract and pulling out the obligations, turning six months of customer emails into the questions you keep answering, drafting the SOP you have been putting off. Available in the Claude apps on paid plans. Give it clear boundaries and review what it produces.
Gemini voice in Gmail, Docs and Keep (Google)
Google has been rolling out natural voice input across Workspace. Speak a five-minute end-of-day brief into Keep and let it come back organised. For anyone who thinks better out loud than on a blank page, this removes the hardest part of documenting anything. Rollout depends on your plan.
Fathom (free meeting notetaker)
Speech-to-text got cheaper again this week (xAI’s new transcription model costs developers about ten cents per hour of audio), and consumer tools ride that wave. Fathom joins your video calls, records, transcribes and hands you a summary with action items. The free tier covers most small teams. Tell the other party it is recording.
HubSpot Breeze invoice prioritisation
If you already run HubSpot, Breeze now ranks your overdue invoices by revenue risk, customer risk and invoice risk, then drafts the follow-up. The ranking is the useful part, because the biggest invoice is rarely the biggest risk. Let it order the list. You decide the tone and the timing.
If you are still deciding where AI fits at all, the pillar guide to implementing AI in a small business walks through the order of operations. Workflow before automation. Always.
Quick win: the 20-minute lock check
You do not need a security consultant to close the gaps this week’s stories exploited. You need twenty minutes. Do these in order.
- Turn on two-factor authentication for email, banking and accounting. Email first, because whoever controls your inbox can reset everything else. Use an authenticator app rather than SMS where the option exists.
- Kill the shared login. If three people sign into Xero, MYOB or your CRM as the same user, you cannot tell who did what or remove one person without locking out the rest. One person, one login, and a password manager.
- List every AI tool that can touch your data. Which ones can read your email, your files, your CRM? Write it down. Remove any connection you set up “to try it” and never used.
- Write a two-line payment rule. Any change to bank details or urgent payment request gets a phone call to a number you already had on file. Not the number in the email. Put it on the wall.
- Update the tools that update themselves. If anyone on your team uses AI coding assistants, confirm they are on this week’s patched versions.
Two of the three companies Gemini got into during Google’s test were breached with credentials left sitting in public code. Nobody had to be clever.
Watch out for
1. The “web of deception” investment scam
ASIC’s latest warning is blunt: a quick Google search is no longer enough to verify an investment. Scammers use generative AI to build a whole ecosystem: a polished platform, fake performance charts, hundreds of five-star reviews and fabricated news articles featuring familiar Australian faces. ASIC removed more than 19,400 online scams in the last financial year, up 182 per cent. Do this: before the business puts money anywhere, verify the licence number on ASIC’s professional register yourself and check the Moneysmart investor alert list. If the details do not match, walk away.
19,400+ online scams taken down by ASIC in FY26, a 182 per cent increase on the previous year. Fake investment platforms alone were up 151 per cent.
2. The voice that sounds like your boss, your supplier or your bank
Voice cloning now runs on consumer apps. The pattern is always the same: urgency, a request to pay or change bank details, and pressure not to check. Australians reported over $157 million lost to scams in the first half of 2026. Do this: the payment rule above. A call back to a known number defeats every version of this, no matter how good the voice is.
3. The add-on that updates itself
Plugin4Shell is a reminder that the add-ons your team installs to work faster inherit every permission that person has. A plugin that quietly updates in the background is a door you did not know was open. Do this: only install AI extensions from the vendor’s official directory, and keep a simple register of what is installed, who owns it and how to switch it off.
Big picture: most owners are not behind, and that is the point
Here is the number that stuck with me this week. A survey of 1,000 Canadian small business owners found three in five say AI has had little to no impact on their business, or they are not sure how it applies. Only nine per cent credit it with real productivity gains. Given a hypothetical $10,000 grant, just seven per cent would spend it on AI.
Three in five small business owners say AI has had little impact or they do not know how it applies. Only 9 per cent report real productivity gains.
That is Canada, not the Illawarra, so treat it as directional rather than gospel. But it matches what I hear from owners in Wollongong every week. The headlines say everyone is automating everything. The workshops say most people have a ChatGPT tab open and no idea what to do with it beyond the odd email.
Both things are true at once. A three-person team can break into one of the most valuable companies on earth in a weekend. And most small businesses have not felt any of it, because capability was never what they were missing. Clear process was. The nine per cent getting real gains almost certainly did the boring work first: wrote down how the job is done, decided who owns it, then handed the repeatable part to a tool.
So if you feel behind, you are not. You are the majority. The question is not which model to pick. It is which one job in your business is repeated often enough, and defined clearly enough, to be worth handing over. If you are in the Illawarra and want a second set of eyes on that, the AI automation for Wollongong businesses page explains how we approach it.
Workflow before automation. Always.
Do the lock check this week. It takes twenty minutes and it is the cheapest insurance you will ever buy. See you next Monday.
Ashley
