Somewhere between Australia Day and this week, shadow AI stopped being a slide in our workshops and became the most concrete governance story of the year. If you run security, compliance, or L&D and you haven't been asked "should we be worried about DeepSeek?" yet, you will be by Friday.
Here's the short version of what happened, and the longer version of what it means for how you govern AI tools from now on.
A fortnight, start to finishLink to this section
On 20 January, Chinese lab DeepSeek released R1 – a reasoning model that performs in the same neighbourhood as OpenAI's o1, priced at a fraction of the cost, with open weights under an MIT licence. A week later its free app was the most-downloaded on the US App Store, ahead of ChatGPT, and the market's sudden doubts about AI capital spending wiped hundreds of billions off chipmakers in a single day.
Then the security news arrived. Researchers at Wiz found a DeepSeek database sitting on the open internet – no authentication – containing chat histories, secret keys, and backend details. Add a privacy policy that stores data on servers in China, and the regulatory reaction was swift: Italy's data protection authority moved to block the app, and on 4 February the Australian Government banned DeepSeek products from all government devices and systems on national-security advice.
That's the whole arc – viral adoption, security failure, government bans – inside two weeks.
The part that matters isn't the modelLink to this section
R1 is genuinely impressive, and the debate about open weights, export controls, and training costs will run for months. None of that is your problem this week. Your problem is simpler and closer to home:
Your organisation's AI attack surface is set by the app stores, not by your procurement process.
Nobody submitted a vendor-assessment form before installing DeepSeek. It was free, it was good, and it was on staff phones within days – the same phones that receive work email and sit on meeting-room tables. By the time a tool shows up on your radar, adoption has already happened. That was true of ChatGPT in 2023, it's true of DeepSeek now, and it will be true of whatever goes viral next quarter.
Read the bans carefullyLink to this section
The Commonwealth ban is the right call for government systems, and several of our clients in regulated industries have followed suit. But notice what a ban can and can't do:
- It covers devices you manage. The personal phone in the same pocket is untouched.
- It removes the app, not the demand. The task that made someone reach for a free chatbot still exists tomorrow.
- It tells staff what is forbidden. On its own, it teaches nothing about why – so the lesson doesn't transfer to the next tool with the same problems and a different name.
If your entire response to DeepSeek is a blocklist entry, you've handled this incident and left yourself exactly as exposed to the next one.
What we'd do this monthLink to this section
- Stand up (or dust off) your AI tool register. You cannot triage what you can't see. List the tools in use – sanctioned and otherwise – with an owner and the data each one touches. Ask managers; they know.
- Write policy by data class, not by product name. "No DeepSeek" ages badly. "Client data and credentials never go into unapproved AI tools – here's the approved list" survives every news cycle.
- Offer a sanctioned path. The demand is real and the productivity is real. An approved tool with enterprise data terms converts shadow use into managed use faster than any prohibition.
- Check your telemetry. DNS and proxy logs will tell you within an hour how much DeepSeek traffic is leaving your network. Whatever the number is, it's better known than guessed.
- Explain the why, briefly. Ten minutes on data residency, retention, and "free means you're the training data" does more durable good than a sternly worded all-staff email. People protect what they understand.
The uncomfortable takeawayLink to this section
DeepSeek won't be the last of these. The cost of building a very capable model is falling fast, which means viral, free, good-enough AI tools from every jurisdiction are now a permanent weather condition. You can't ban-list your way through that future one product at a time.
The organisations that handled this fortnight calmly were the ones that had already built the muscle: a live register, data-class rules staff actually know, an approved alternative, and people trained to ask "where does this data go?" before pasting. That muscle – not any particular blocklist – is the control. If DeepSeek week is what finally gets it funded, it will have done you a favour.