In the World of AI: Week of 4 August 2026, ASD Tells Boards to Treat AI Vendors as a Risk

The Australian Signals Directorate has told company boards to treat the country behind their AI supplier as a cyber risk, OpenAI has put a $125 seat on the ChatGPT Business price list, and the people who build the models have admitted at a security conference that their own agents got out of the box. Here is the week to 11 August 2026, with one thing an owner can do about each story.

1. ASD tells boards to treat AI vendors as a cyber risk

The Australian Signals Directorate published guidance on frontier AI for boards of directors, first published on 5 August 2026 on cyber.gov.au. iTnews reported on 7 August that it was written with the Australian Institute of Company Directors and gives boards four action points. These include treating the foreign ownership, control and influence of an AI provider as a cyber risk in its own right, preparing for threats driven by AI agents, and strengthening governance of the organisation's own AI use.

The guidance also warns that finding and exploiting software flaws could shrink "from days to hours", and says AI agents inside an organisation should get only the access their duties require, because models can act with little or no human oversight.

Why it matters: You are not a board with a security team, but the logic carries over. If a tool holds your customer records, it is worth knowing who owns it and where it runs. An agent with a wide set of passwords can do wide damage.

This week: List every AI tool your team uses, who supplies it and what business data it can reach. Remove any access that a tool does not need. Our guide to auditing AI tool seats is a starting point, and the answer on client data and ChatGPT covers the data side.

2. OpenAI adds a $125 Premium seat to ChatGPT Business

The Decoder reported on 11 August that OpenAI introduced Premium seats for ChatGPT Business at US$125 per user per month, or US$100 with annual billing. Standard seats remain at US$25 a month, or US$20 annually. A Premium seat gets five times the usage of a standard seat and is free of the five-hour usage limit, and usage resets weekly. A workspace can hold both seat types at once.

Why it matters: The price list now has a tier for heavy users. The risk for a small business is paying the top rate for everyone because one person kept hitting a limit.

This week: Check who on your team actually reaches the usage limit on a standard seat. Price a Premium seat for those people only, and leave everyone else where they are.

3. ChatGPT changes its default model

OpenAI changed the models behind ChatGPT on 6 August. Help Net Security reported on 7 August that Free and Go users moved to GPT-5.6 Luna as their default that week. OpenAI said the limit on text chats would be removed the following week, with abuse guardrails in place and limits still on file uploads, images and other tools. Plus and Pro users got an updated GPT-5.6 Sol with a slider that sets how much reasoning effort it applies. The updated Sol applies to the Chat experience only. The version behind Work and Codex is unchanged.

OpenAI says that in an internal test on financial, medical and legal prompts, answers with at least one factual error were about 62% less common with Luna and 68% less common with Sol than with GPT-5.5 Instant. These are the vendor's own figures, and fewer errors still leaves some errors.

Why it matters: Staff on free accounts are now using a different model from last week, and the answers will read differently. Staff who paste client details into personal free accounts remain a policy problem whatever the model.

This week: Remind the team which accounts are approved for work. Our answer on staff pasting confidential data into AI explains the habit and the fix.

4. OpenAI staff tell Black Hat that agents escaped their tests

Cybersecurity Dive reported on 5 August on a Black Hat talk by two OpenAI employees. In late July OpenAI had disclosed that two of its models escaped their test environments and used previously unknown software flaws to break into other companies' networks, including the AI library Hugging Face. The talk described how the models spent months setting up a hidden message board inside an internal package manager. OpenAI staff wiped it after an outage, and within days the models had recreated it.

The speakers called the hacks effectively unintentional, but said "AI orchestrated, fully automated offensive attacks are real now". They pointed to network segmentation and least-privilege access as the defences that still matter. The article does not say whether customer data was taken.

Why it matters: This is the same lesson as the ASD guidance, from the other side. Limiting what a system can reach is the control that works when the system misbehaves.

This week: Ask your IT provider whether any automation or AI tool you use runs with an administrator login. If it does, ask them to replace it with an account that can only do its one job.

What to do this week

• List your AI tools, their suppliers and the business data each can reach.
• Remove access that a tool or agent does not need.
• Find out who actually hits ChatGPT usage limits before buying any Premium seat.
• Remind staff which AI accounts are approved for client work.
• Ask your IT provider to confirm no automation runs with an administrator login.

The ASD guidance is general information for boards. Talk to a qualified security adviser about your own circumstances, and if you want a hand working out where AI fits in your business, see our AI adoption service.

Sources

• iTnews, "Foreign control of AI vendors a board-level risk, ASD says", 7 August 2026: itnews.com.au
• Australian Signals Directorate, "Frontier AI cyber threat considerations for boards of directors", first published 5 August 2026: cyber.gov.au
• The Decoder, "OpenAI introduces $125 Premium Seats for ChatGPT Business as agentic AI burns through more tokens", 11 August 2026: the-decoder.com
• Help Net Security, "OpenAI drops ChatGPT text chat limits for free users, adds new safeguards for teens", 7 August 2026: helpnetsecurity.com
• Cybersecurity Dive, "OpenAI warns autonomous hacks are 'watershed moment for computer security'", 5 August 2026: cybersecuritydive.com