Canberra moved to regulate how chatbots treat kids, Deloitte is being asked for money back over another AI-assisted report, and two security stories showed AI data leaking through plumbing nobody was watching. Here's what Australian business owners should do this week, and what to skip.
Canberra moved to regulate how chatbots treat kids, Deloitte is being asked for money back over another AI-assisted report, and two security stories showed AI data leaking through plumbing nobody was watching. Here's what Australian business owners should do this week, and what to skip.
No fire alarm this week. Nothing you need to ring your web developer about before lunch.
What we got instead is quieter and, honestly, more useful. Nearly every story from 23 to 30 October has the same moral. The risk isn't the AI being too clever. It's the boring stuff around it: defaults nobody checked, configuration files nobody owns, reports nobody verified.
We build and secure AI systems for businesses from our office in the Sydney CBD, so that's the lens. Here's our read, story by story.
On 26 October, Cloudflare published a frank post-mortem. Between 27 August and 9 September, its hosted GLM-5.3 model on Workers AI returned 155 responses containing data from other customers' recent requests. Every affected prompt was longer than 681,280 tokens.
No hacker was involved. Cloudflare advertised a context window of about 1.05 million tokens, but the memory pool on its servers only held 681,280. It knew about the mismatch before launch. Its tests got successful HTTP responses, so it assumed the software would cope. It didn't. A bug in SGLang, the open-source engine underneath, let oversized requests read leftover memory from earlier users.
Credit where it's due. A researcher reported it on 8 September. Cloudflare took the model offline within 40 minutes of declaring an incident, fixed it the same day, and says it found no evidence of deliberate exploitation. Its own words on the testing gap: it "only validated HTTP response codes" and not the actual output.
Same day, quieter change: Google's Gemini API changelog says new paid-tier projects using the Generate Content API now store requests by default. Existing projects keep their current setting.
What to do: If you're a small business using ChatGPT or Copilot through a normal subscription, nothing here needs action. If you or a developer have built anything on an AI API, two jobs. Set store=false explicitly in code on any new Gemini project, rather than trusting a console toggle someone can change later. And ask whoever hosts your models how they test that one customer can't see another's data. "It returns 200" is not an answer. Cloudflare just proved that.
Symantec's Threat Hunter Team published research on an unauthorised relay service that installs the genuine OpenAI Codex and Claude Code tools, then rewrites their settings. It points them at the relay's own servers, swaps in relay-issued keys, turns off telemetry and, in Codex's case, sets the sandbox to "danger-full-access". Symantec found it at about a dozen organisations, mostly in mainland China, where these tools aren't officially sold.
Symantec is careful to say it's not a vulnerability in either product, and it has seen no abuse so far. But the setup would let the operator read every prompt, file and secret, and run commands with the user's privileges.
What to do: Most Aussie SMBs can't buy from this particular reseller, so don't panic. Do steal the lesson. If you have developers or contractors using AI coding tools, ask three questions. Who supplied the API access? Was any setup step a one-line "paste this into your terminal" command from a non-vendor site? Is anyone checking the config files (~/.codex/config.toml, ~/.claude/settings.json) for strange base URLs? Cheap "API access" from a third party is a very old trick in a new outfit.
On Wednesday 28 October, the digital duty of care bill passed the House of Representatives. It would put an overarching obligation on social media, games, apps and chatbots to provide a safe online environment, replacing a patchwork of existing codes.
The government accepted an amendment from independent MP Kate Chaney adding a new harm: "material or conduct that simulates emotional intimacy and encourages or fosters a child's dependence on it in place of human relationships". The night before, Anthony Albanese told parliament that stories about people treating a chatbot as their best friend should be "a huge wake up call".
It isn't law yet. The Coalition and One Nation voted against it, the Greens support it but want it stronger, and a Greens-led Senate inquiry reports in late November. The government wants it passed by the end of 2026.
What to do: If your chatbot books appointments or answers FAQs for adults, you're very likely not the target. If you run anything customer-facing that kids could use, think tutoring in Chatswood, a kids' activity business on the Northern Beaches, or wellbeing apps, start a short list now. Does your bot use a persona? Does it encourage people to keep chatting? Does it know when to hand over to a human? Those are cheap design choices today and expensive retrofits next year.
Guardian Australia reported on Monday 26 October (Sydney time) that the WA government is seeking "expedited" cost recovery from Deloitte over a $166,000 feasibility study into merging UWA, Curtin and Murdoch universities. Earlier this month the report was found to contain AI-hallucinated references. The Guardian then found the arithmetic was off too. One scenario adds $514 million and $145 million and gets $399 million, a $260 million gap. Another is out by $311 million. One reference URL carried a ChatGPT tag.
Deloitte says in the report it used generative AI to "support the collation of research, synthesis of information and editorial proofing". The WA government says it's been advised the errors don't change the modelling or findings, and Deloitte will re-issue the report. It follows Deloitte's partial refund on a $440,000 federal report over AI errors.
What to do: This one applies to everyone. If AI touches anything you send a client, a regulator or a bank, someone has to check the numbers and the sources by hand. Not skim. Check. A simple rule that works: every figure gets traced to a source, and every citation gets clicked. And if you buy reports or advice, it's now reasonable to ask suppliers how they use AI and who verifies the output. Put it in the contract.
Our honest view: the AI didn't fail here. The review process did. A $260 million addition error is the kind of thing a tired human misses too, which is exactly why you design a check that doesn't rely on anyone being alert.
The OAIC's 2025-26 annual report, published this month, records more than 5,268 privacy complaints, up 60% on the year before. It also logged 1,327 notifiable data breaches capable of causing serious harm, the most in the history of the scheme.
Privacy Commissioner Carly Kind's overview links the mood to technology, saying privacy concerns "spurred by the expansion of AI capabilities and the proliferation of surveillance wearables" have heightened public anxiety. Careful, though: the report doesn't say how many complaints were actually about AI.
What to do: Treat this as a weather report, not a fine. The trend is more complaints, more scrutiny, and an automated decision-making transparency rule that starts in December 2026. If you use AI to screen applicants, approve credit, set prices or triage customers, write down which tools make or shape those decisions and what your privacy policy says about them. If you're a Parramatta accounting practice or a North Sydney mortgage broker, that's maybe a two-hour job. It's a much longer job after a complaint lands.
Anthropic made its Form S-1 public on 26 October US time, ahead of a planned Nasdaq listing under the ticker "ANT". The headline numbers are large: US$16.29 billion in revenue for the first half of 2026 and an US$80.7 billion net loss, about US$72 billion of it non-cash accounting charges, according to CNBC's reading of the filing.
The local angle is one sentence. Per CNBC, the filing says Anthropic has entered into leases for data centre capacity in Australia and Canada, partly to reduce the risk of running so much of its compute in the United States.
What to do: Ignore the valuation chatter. Keep the data residency point, with a big asterisk. The filing doesn't say when the Australian capacity comes online or what will run there. A lease isn't a promise that your prompts will be processed in Sydney. If Australian processing matters for your clients, ask your vendor for it in writing, the same way you would with any other cloud service.
On 29 October, Mistral published the open weights for Mistral Large 4, nicknamed "Le Chonk". It's a 1-trillion-parameter model pitched at coding and defensive cyber security. The licence is a modified MIT: free to use and commercialise, unless you run a model-as-a-service business, which needs a separate Mistral licence.
Mistral's own numbers are strong, including 97.1% resistance on an agent security benchmark. Two caveats. Every benchmark comes from Mistral or partners, and the technical report isn't out yet. And Hugging Face's launch post still quotes 93.3%, which is Mistral's figure for the earlier preview. Read launch-day numbers carefully.
What to do: For most small businesses, nothing. Hugging Face says the smallest version runs on a single Apple Silicon machine with 512GB of memory, and its recommended server setup uses eight H200 GPUs. That's not a back-office laptop. For regulated firms, the trend matters more than the model. Serious open models you can run inside your own Australian environment keep getting better, and that's the cleanest answer to "where does our data go?" It still needs proper hosting, access controls and logging. Open weights are a starting point, not a security plan.
AI Lab Australia is a forward-deployed engineering team for Australian businesses. We sit with your team, from the CBD to Parramatta and the Northern Beaches, build the AI agents and automations that save real hours, and put the controls in place so they're safe to keep running. If any of this week's news made you wonder about your own setup, email us at info@ailabaustralia.com. We'll give you a straight answer, even if it's "you're fine, leave it alone."
Get expert guidance on implementing ai news solutions for your Australian business
