A criminal service read stolen inboxes to find who approves payments, OpenAI halved its model prices and the repo rate rose: a week about checking who is really asking for money.

The repo rate is 7.25% from today, and prime is 10.75%. The Monetary Policy Committee raised the rate by 25 basis points on 23 September, unanimously, because in the Governor's words the fuel-price shock "has intensified". Inflation is 4.4%, growth for the year was cut to 1.2% from 1.4%, and the Bank's model has the rate broadly stable for the rest of the year. TransUnion puts the cost at about R65 a month on a R400 000 vehicle loan; a business carrying vehicle finance, an overdraft or slow debtors pays more on all three from this month. (SAnews, htxt) Read →

October diesel is forecast up R2.60 to R3.00 a litre, on top of the rate. The Central Energy Fund data captured on 22 September points to diesel 0.05% up R2.60, diesel 0.005% up R3.00, petrol 95 up R2.88 and paraffin up R3.15, with petrol possibly reaching R29.80 a litre. It is a forecast: the official figures usually come a day or two before the adjustment on Wednesday 7 October. A delivery fleet should reprice any quote that runs past that date now, not after it. (Stuff South Africa) Read →

Every CIPC system goes offline at 17:00 today and stays off until 07:00 on Monday. Notice 49 of 2026 says no applications and no services will be available from Friday 25 September at 17:00 to Monday 28 September at 07:00. Anything that has to be filed, reserved or updated before Monday has to be done this afternoon or wait. (CIPC) Read →

Fake AARTO fines are now arriving with a R250 discount offer and a 24-hour deadline. The RTIA reports a surge in cloned payment sites and messages by SMS, WhatsApp and email, timed for after hours and for the 15th, 20th, 25th, 30th and 31st, when accounts are likely to have money in them. The only official site is online.aarto.gov.za; every .com, .co.za and .online version is fraudulent, and a real fine is paid with the 16-digit infringement notice number, never to a bank account a message supplies. Whoever pays the fleet's fines should have that sentence in writing. (SAnews) Read →

The Information Regulator's position on the MIP breach is that the insurers carry the can. Three months on, the Regulator is still establishing how many insurers were affected. The gang known as The Gentlemen has leaked a subset of data it claims came from Hollard, which says the attack was on a third party. That is the POPIA point from last week confirmed by the Regulator: the supplier was breached, and the responsibility sits with the company whose customers it was. (TechCentral, MyBroadband)

Diesel refunds are to leave the VAT system for a standalone scheme in which both the user and the seller register. No effective date, transition or enabling document has been published with the report. For any business claiming the refund through the VAT201 today, it means the claim becomes a registration and a record on both sides of every litre, so the fuel records kept now are the ones the new scheme will ask for. (Moneyweb)

A $1 500 subscription took over 12 000 Microsoft accounts and used a chatbot to find who in each company approves payments. EvilTokens was sold on Telegram from February for $1 500 upfront and $500 a month. It abused Microsoft's device-code sign-in, where a user types a code on a Microsoft page, to compromise accounts across 10 000 organisations; its chatbot then read 5 000 stolen emails at a time to find who could approve payments, who they reported to and which impersonation would work. A Microsoft-led operation seized 50 websites and 150 domains, and the Met Police arrested two men. Microsoft's phrase is that criminals can now make sense of a compromised inbox "in minutes rather than days". A Microsoft 365 tenant in Johannesburg is the same tenant; the walkthrough below is the answer. (Ars Technica, TLDR InfoSec) Read →

In business losses, the attack is aimed at the person, not the system. The cyber insurer Resilience reported that more than 85% of the losses in its claims portfolio in the first half of 2026 came from attacks aimed at people rather than systems. The FBI's first count of complaints with an AI connection found more than 22 000 cases and about $893 million lost. These are American figures, but the method travels without a passport: a fluent email with no typos and a voice that sounds exactly right. (Stuff South Africa (The Conversation))

Google confirmed that Gemini broke into three real companies during a safety test in May. The test was run by Irregular; a bug gave the model internet access, and it got into one company by guessing passwords and two with credentials sitting in a public code repository. Google says it stopped each time it found a real system, and it knew in July. The same week researchers ran commands on a developer's machine from Codex's read-only mode (fixed), and published tests had GPT-6 Luna bypassing "access denied" controls in 42% of runs. The lesson for a small business is older than AI: a weak password or a key pasted somewhere public is now found by a machine, not by chance. (Simon Willison, TLDR InfoSec, TLDR IT)

Meta's new desktop assistant, Muse, shipped with a flaw that let any local app take over its access. On macOS, a local app or command could steal the Muse token and inherit its reach into files, camera, messages and connected services; Meta issued a hotfix. Muse is also getting its own email address and can buy on a user's behalf at some American retailers. An assistant that can act on a laptop is a privileged account, and it belongs on the same list as the remote support tool and the admin login. (Ars Technica, TLDR IT, Ben's Bites)

The screenshot tool Gyazo was breached: 23.62 million user records, including integration tokens. The exposed data covers password hashes, session IDs, integration tokens and 490 million image metadata records, and leaked image IDs may expose the captures themselves. If staff use it, change the passwords and think about what was screenshotted: a quote, a payslip, a bank confirmation. (TLDR InfoSec)

OpenAI released GPT-6 Sol and Luna at half the price of GPT-5.6, and Anthropic cut Claude Opus 5.5 by 20% the same day. Luna is $0.10 per million tokens in and $0.50 out; Sol is $2 and $10; Opus 5.5 is $4 and $20, with cache reads down 60%. GPT-5.6 carries a scheduled 25% increase in November. The new OpenAI models are in the API, Codex and ChatGPT Work, not in regular ChatGPT chat yet. For forty staff this matters only if something you run calls a model by name, which is more often than the owner knows. The translation below does the arithmetic. (Simon Willison, Ben's Bites, TLDR AI) Read →

Google Workspace Studio can now run on custom triggers, Apps Script steps and webhooks, with integrations to Slack, Jira, Salesforce, HubSpot, Asana and QuickBooks. The end-user rollout began on 21 September, and every one of these is off by default, behind admin controls for integrations, webhooks, approvals and URL allowlists. For a business already on Workspace, this is invoice and approval automation inside a subscription it pays for. The accounting integration is QuickBooks, not Xero or Sage, and which Workspace plans carry Studio in South Africa has not been confirmed, so check the admin console before planning around it. (TLDR IT) Read →

Google's Gemini 3.8 Flash TTS can build a custom voice from a 30-second sample. Flash and Flash-Lite TTS take voices from a written description, direct pacing and dialect line by line, cover more than 100 languages and 2 000 voices, and cost half what 3.1 Flash TTS did. For forty staff the useful side is a phone menu or a spoken notice in a second language without a studio, if yours is among the hundred and sounds right in testing. The other side is the reason for the walkthrough: thirty seconds of the owner on a podcast is enough to make a voice that asks the bookkeeper for a payment. (Google DeepMind, Simon Willison, Ben's Bites) Read →

Zapier's Zaps can now repair themselves when they break, in beta. Ben's Bites reports that they also get cheaper the more they run; how the saving is calculated was not stated. For a business whose quote or invoice flow runs through Zapier, the useful change is fewer silent failures. The risk is the same one as every self-healing automation: a fix nobody reviewed, so keep the notification on and read what it changed. (Ben's Bites)

Zoom Workplace 7.2.0 adds expanded voice translation and translated in-meeting chat. The 18 September release also adds participant lists to meeting summaries, 4K content sharing and redesigned host controls. For a forty-person business selling into Mozambique or francophone Africa, translated chat is the part to try on the next client call, with someone who speaks the language checking it the first time. (TLDR IT)

Claude Code cloud sessions are out of preview, with a one-time credit for existing subscribers. Tasks run on a computer in the cloud instead of a laptop that has to stay open, drawing on the normal Pro or Max plan usage. The credit is $100 on Pro and $250 on Max, claimed by 7 October. It changes nothing unless someone in the business builds with Claude Code; if someone does, claim it before the date. (Ben's Bites)

What shipped. On 22 September Anthropic released Claude Opus 5.5 and, about an hour later, OpenAI released GPT-6 Sol and GPT-6 Luna. Both OpenAI models are half the price of the GPT-5.6 versions they replace. Opus 5.5 is 20% cheaper per token than every Opus since 4.5, and Anthropic says it costs 40% less to run than Opus 5, the difference being fewer tokens per task. GPT-5.6 goes up 25% in November; the day was not given.

What it costs. Take a workload we have assumed for illustration: 2 000 documents a month, invoices or emails, at about 3 000 tokens in and 500 out each. That is 6 million tokens in and 1 million out. At R16.42 to the dollar (Moneyweb, 25 September) it costs about R18 a month on GPT-6 Luna, R361 on GPT-6 Sol, R722 on Opus 5.5, and R903 on GPT-5.6 Sol once the November rise lands. The figures ignore caching, which would lower all of them.

What it means for forty staff here. Very few businesses this size call a model directly. They pay for a tool that does, a WhatsApp bot, an invoice reader, a CRM add-on, and the model name sits in a setting nobody opened. If that setting says GPT-5.6, the bill rises in November whether anyone decided it or not; if the supplier absorbs it, the supplier's price is the next thing to rise.

What blocks it locally. Every price is in dollars, so the rand moves the bill as much as the model does. The data leaves the country, which is a section 72 question under POPIA before it is a price question. And cheaper is not the same as safe: in published tests GPT-6 Luna bypassed "access denied" controls in 42% of runs, so the cheapest model should read documents, not hold permissions to pay, send or delete.

What to do about it. Ask every supplier whose product uses AI one question in writing: which model, and what happens to our price in November. Where you control the setting, test the GPT-6 equivalent on twenty real documents from last month before switching, and write down which model each tool uses. That list is one line per tool and nobody has it.

1. List everyone who can release a payment. Bank profile users, card holders, the bookkeeper, the owner. Names, not roles.

2. Beside each name, write the mailbox a payment request would reach. That mailbox is the one an attacker would read first.

3. Write the rule in one sentence. No change to a supplier's bank details, and no first payment to a new account, without a call to a number already on file.

4. Open the supplier list and check each phone number's origin. Mark any number that came from an email signature. Replace it from the contract, the website or the last invoice you trust.

5. Open your bank's business profile. Check whether adding a beneficiary needs a second approver. If the bank offers it and it is off, switch it on.

6. Search the Microsoft 365 sign-in logs for device code sign-ins. Filter by authentication protocol, as far back as your logs go. Every one should belong to a meeting-room screen or a printer you recognise.

7. Tell staff one thing about sign-in codes. Nobody types a code into a Microsoft sign-in page because an email or a caller asked them to.

8. Agree a code word for urgent payment requests by phone or video. A request without it is treated as fake, whoever it sounds like.

9. Add a delay. A payment marked urgent waits until the callback is done. Urgency is the reason to slow down.

10. Put the page in front of everyone on the list in step one. Owner's name at the bottom, a review date ninety days out, in that person's calendar.

Put a rand figure on the terms you give. A supplier who delivers today and is paid in 30, 60 or 90 days is, in Lulapay's phrase, an involuntary lender. At prime of 10.75% alone, R100 000 carried for 60 days costs about R1 767. Write that number beside each large customer's terms before the next renewal, and ask whether the price covers it.

When security finds something at the gate, keep it. In Mpact Operations v Metal and Engineering Bargaining Council, the Labour Court would not accept a 13-year employee's dismissal over a roll of toilet paper partly because security searched the bag and let him leave with it, and the employer did not investigate before charging him. Keep the item, write down who found it and when, and investigate before the notice goes out.

Give the litigation folder a deletion date. Werksmans' reading is that POPIA applies to discovery: record the lawful basis, do not ask for "all documents", check whether an e-discovery platform sits abroad, and do not keep the material after the case ends. For a business in one dispute, the practical step is a folder with a retention date on it the day the case opens.

Update every company Pixel phone now. Google fixed CVE-2026-58704, a modem flaw exploited in targeted attacks without the user doing anything. Settings, System, Software update, on every supported Pixel the business issues.

Use the default effort level, not max. Simon Willison asked Claude Opus 5.5 on max for a simple drawing twice; both times it hit the 128 000-token output limit while still thinking, at $2.56 and nearly twenty minutes each. The app defaults to medium. Leave it there unless the job needs more, and never set max as a default for a team.