AI policy people are arguing again about how open powerful models should be, after an Anthropic report flagged the open model GLM-5.3 as a major cyberattack risk. For a small IT team, this is a reminder that open-source AI tools may be great for cost and control, but you should treat anything security related as you would unvetted code from the internet.
Experts in the same debate also point out that open models help with transparency, local hosting, and custom use cases. If you are testing open models on-prem or in your own cloud, keep them in a clear sandbox and avoid plugging them straight into production workflows that touch client data or identity systems.
AI training data and dead startups
MIT Tech Review reports that AI labs are buying data from failed startups, with one founder earning around $25,000 for his code base and more for user data, as labs hunt for fresh material to train models. For you, this raises a quiet risk that old SaaS tools or side projects used by your staff might have their data sold on, so it is worth tracking which services hold company data and how they handle shutdown or asset sales.
The same piece notes that some deals include chat logs, internal documents, and other sensitive records from these dead platforms. IT should tighten offboarding for SaaS, including closing unused accounts and asking vendors for written details on what happens to stored data if the business is sold or wound up.
Stop trusting AI to say no for you
Another analysis argues we are putting too much faith in AI systems to refuse bad requests, even though refusal behaviour is easy to work around. If you are piloting AI for support or decision help, do not rely on the model to block risky actions, keep your own policy checks, approvals, and technical controls in place.
Need IT support that just works? Email hello@icthero.co.uk or talk to us.