The Information Regulator wants fines with no grace period. Four AI assistants down in one afternoon. UberX switched off on 1 September. A second subscription is not a fallback.

The Information Regulator wants the grace period gone, and has started writing to companies nobody complained about. At an enforcement briefing on 31 August it said it intends to ask Parliament for immediate fines on a finding of non-compliance, rather than the current arrangement where fixing the problem after an enforcement notice means no fine. It has also begun asking responsible parties to demonstrate compliance before any complaint exists. Fines to date: R5m against the Department of Justice and R5m against Basic Education, both contested, R100 000 against the IEC and about the same against Lancet Laboratories, both paid, and R500 000 against Bloubergstrand Municipality, reduced by a court to R250 000. (Werksmans, ITWeb)

The two diesel grades moved by different amounts on 2 September, and 26.8 cents of the increase is not the oil price. Diesel 0.05% went up 294c a litre and 0.005% went up 315c, so a rate card that does not record which grade the fleet takes is wrong by 21c a litre before anyone notices. Inside the increase sit a 21.9c slate levy and a 4.9c adjustment for forecourt wages, neither of which behaves like Brent. (Stuff South Africa, Daily Maverick, Bizcommunity Logistics)

ChatGPT, Claude, Grok and Gemini all had trouble in South Africa on the same afternoon. Users reported problems with all four on Thursday 3 September. No cause, no duration and no vendor confirmation has been published, so treat the times as unknown and the correlation as the whole story. Keeping a second subscription is the standard hedge against one model going quiet, and on Thursday it was not a hedge. (MyBroadband)

Uber discontinued UberX in South Africa on 1 September, and left Nigeria entirely the next day. The Nigerian wind-down ends twelve years there and comes with roughly 3 000 layoffs globally. It is the third African exit in about a year, after Tanzania in January and Cote d'Ivoire in September 2025. If a meaningful share of your work arrives through somebody else's app, the customer record, the price history and the job history are the three things that decide whether a switch is a fortnight or a rebuild. (TechCabal)

Shoprite now owns 51% of the payments platform thousands of spaza tills run on. The group closed an initial majority stake in R&A Cellular, giving it a route into financial and value-added services in informal and semi-formal retail. Read past the retail angle: this is a business's operating software changing hands to a competitor, and the question it raises for everyone else is who ends up holding your transaction history when your supplier's shareholder register moves. (TechCentral, ITWeb, Bizcommunity ICT)

Sugar import duty rose about 44% and ITAC is preparing to charge for permits it used to issue free. The duty moved on Friday 28 August from 483.72c a kilogram. The permit fee regulations arrived with no fee schedule, no effective date and no list of affected permits. Both land on a landed-cost model, and in most distributors the duty rate is a number somebody typed into a spreadsheet once. (Freight News)

A pharmaceutical distributor lost 284 million records because someone answered the phone. McKesson detected an intrusion on 25 August. The attackers telephoned staff, took over single sign-on accounts and pulled roughly a terabyte out of Salesforce and Snowflake over four days, then demanded just over $55 million, which was not paid. Nothing in that chain required a clever exploit, and the helpdesk that resets a password is the same shape in a business of forty. (TLDR InfoSec)

The window between a public proof of concept and an attack is now about two days. CrowdStrike reports that 88% of the vulnerabilities it observed between January and June were actively exploited within 48 hours of a proof of concept becoming public, with the fastest groups under 24 hours. A monthly patch cycle was built for a world where that window was weeks. (TLDR InfoSec)

Microsoft 365 spent a Monday down, and the cause was a configuration change to authentication. It began with Exchange Online on 31 August and spread to Teams, OneDrive, SharePoint, Purview and Defender XDR before mitigation reached the affected infrastructure. Authentication is the piece everything else is downstream of, which is why the failure did not stay in email. (TLDR IT)

OpenAI gave Cursor's customers ten weeks' notice that the model underneath their tool goes away. Access to OpenAI's coding models ends on 12 November. Nobody bought a model. They bought a tool, and the engine inside it was removed by a decision two companies made about each other. The lesson travels to any subscription where the capability you rely on is somebody else's supplier relationship. (TLDR IT)

Over 400 software packages were poisoned in under four hours, and every one carried a valid signature. The attackers hijacked a maintainer's legitimate signed release pipeline, so the usual supply chain check confirmed the origin was genuine while the code was not. The same worm family has been going after authentication tokens for AI coding tools. If your systems were built by an agency, ask whether they pin versions. (TLDR InfoSec)

The European Commission has designated ChatGPT as a search engine under the Digital Services Act. Reddit and Roblox were classified as very large platforms in the same decision, and all three have four months to meet obligations on systemic risk assessment, mitigation, transparency and independent audit. Europe regulating an assistant as a search engine is the clearest signal yet that what it says about your business is treated as a published result. (TLDR IT)

Atlassian put a spending limit on its AI, and set a date for the bill. Usage-based pricing expands across Rovo credits, automation steps and autonomous agent resolutions, with billing starting 3 December. Admins get centralised usage visibility, alerts and spending limits. The cap is the part that matters: an agent that can call a model in a loop is an unbounded invoice until somebody bounds it, and a ceiling is the only control that works before the month ends rather than after it. (TLDR IT)

Google Workspace Studio can now move a file, answer an email and post a message, with an approval gate on each. New automation steps copy or move Drive files, reply to Gmail threads and post replies in Google Chat. An admin can disable individual actions and require user approval where a workflow could share data outside the organisation. It is available on Workspace, it removes typing, and the approval control is the setting to find before anyone builds the first flow. (TLDR IT)

Chrome halves its release interval from 8 September. Stable moves from four-week to two-week releases beginning with Chrome 153. The eight-week Extended Stable channel for enterprise admins is unchanged. For a business without an IT department this is free and automatic, and it is the single cheapest security improvement in this section. For one with a managed fleet it doubles the testing cadence, which is a diary problem somebody now owns. (TLDR IT)

Okta can now find the AI agents nobody told you about. Its identity posture product discovers shadow AI agents, MCP servers, OAuth grants and agents built inside Copilot Studio and Salesforce Agentforce. If you already run Okta this is an inventory you could not previously produce. If you do not, the useful part is the question: nobody in a business this size can currently list which tools hold a live login to which system. (TLDR IT)

Cloudflare shipped bot detection that retrains itself, on the plan you are probably not on. Adaptive Intelligence continuously retrains on live traffic and delays its reactions so attackers cannot isolate which defence caught them, which is aimed at slow credential stuffing spread across residential proxies. Be clear about the reach: the continuous retraining switch is on the Enterprise tier, not the plan a forty-person business buys, and the disposable rule generation is still in development. (TLDR InfoSec)

Google released Gemini 3.8 Flash at the same introductory price as the model it replaces. Coding, agentic and multi-step reasoning all improve with no price move, which is the rarer of the two ways a model gets cheaper. There is also a Flash Cyber variant for vulnerability detection and automated patching, and that one you cannot buy: it is behind a restricted defender programme, so treat it as a direction rather than a purchase. (TLDR AI)

Anthropic announced monitoring that keeps the logs in your own cloud account. Enterprise Frontier Safeguards pairs zero data retention with automated detection of serious misuse, storing activity logs in the customer's own AWS, Azure or Google Cloud under customer-managed keys, with flags going to the customer's team rather than to Anthropic staff. It is a phased rollout later this year and it is an enterprise arrangement, so it is not something a forty-person business switches on. The idea underneath it is portable and free: monitoring you cannot read is somebody else's evidence. (TLDR IT, TLDR InfoSec)

What happened. Anthropic released Claude Fable 5.1 on 1 September. The standard input and output prices did not move. One line on the rate card did: the price of reading context the model has already been given fell from $1 to $0.25 per million tokens, a 75% cut. Anthropic says that lowers effective cost by about 25% on a typical workload and up to 45% on a heavily agentic one.

What it costs. Those figures are in dollars because they are dollar prices, and no sourced rand rate exists for the week, so converting them would invent precision nobody has. What they turn on is how much of your prompt repeats. For comparison inside the same rate card, Sonnet 5 is $2 per million input tokens and $10 per million output; Fable 5.1 is five times that at standard rates. The earlier 50% discount on Fable 5 ended on 19 July and does not carry over.

What it means for forty staff here. Most business use of a model sends the same block of words every single time: the rate card, the returns policy, the product manual, the tone instructions. That block is now billed at a quarter of what it was on every reuse. It is the one AI cost lever a small buyer can actually pull, because it does not require a bigger contract or a different model, only that somebody structured the prompt so the fixed part sits still.

What blocks it locally. Two things, and neither is the technology. The plan gates the model: a free account cannot select it at all, and Pro and standard Team subscribers need usage credits on top of the subscription before they can use it. And funding the account is a card problem, since being on the supported country list confirms the service works here without guaranteeing your payment method does.

What to do about it. Ask three questions of anyone quoting you AI work, and ask them in writing. Which model, exactly. Do you cache the fixed part of the prompt, or is the price list sent fresh every time. And what reasoning effort do you run at, given the model always reasons, defaults to high, and cannot be switched off. The third one moves both the bill and the wait, and it is a setting, not an argument.

1. List the processes that stop when a vendor stops. Name the process, not the tool. "Invoices go out", not "Xero".

2. Write the vendor beside each one. One name per row. Where two vendors have to work for the process to run, write both.

3. Mark every row where the plan is a second subscription. Those are the rows Thursday tested. Four assistants were down here at once.

4. Rank the list by what a day of it costs you. Money first, then customer promises, then internal inconvenience. Take the top three.

5. For each of those three, write the manual step. One sentence, present tense, addressed to the person who does it. "Phone the customer with the total off the job sheet."

6. Put the vendor's status page link in the row. Find it now. Nobody searches for a status page while the phone is ringing.

7. Set the switch threshold in minutes. Write a number. Twenty minutes down, and the manual step starts.

8. Name the person who calls it. One name per row, not a department. A fallback nobody is allowed to start is a document.

9. Send the three sentences to the three people today. A fallback somebody reads for the first time during an outage is not a fallback.

Set the spending limit before you switch the agent on. Atlassian starts billing usage on 3 December and is shipping alerts and spending caps alongside it. Whichever tool you use, find its ceiling and set it now. A cap set after the invoice arrives is a complaint, not a control.

Require approval on any automation that can send data outside the business. Google Workspace Studio now lets an admin disable individual actions and demand user approval for exactly those steps. The setting exists because the failure mode is real: an automation quietly emailing a file to an external address is the same thing as a leak with a schedule.

Confirm a password reset on a number already on the employee's file. Never the number the caller gives you. A distributor lost 284 million records this week to attackers who telephoned staff and talked their way into single sign-on. The person who resets passwords in a forty-person business is usually also the person who cannot verify who is calling.

Move patching from monthly to inside two days on anything internet-facing. Eighty-eight per cent of observed vulnerabilities were being exploited within 48 hours of a public proof of concept in the first half of this year. The firewall, the VPN, the mail gateway and the camera recorder are the four that face outward and the four nobody updates.

Put a last-verified date on every stored phone number. Kenya's rules on deactivating and recycling numbers take effect on 19 September; South Africa has no equivalent, so a number in your database silently stops belonging to your customer and nothing tells you. Every OTP, delivery notice and account alert you send trusts that field.