Machine Learning Security

I just read an interesting blog post about ML security recommended by Bruce Schneier [1].

This approach of having 2 AI systems where one processes user input and the second performs actions on quarantined data is good and solves some real problems. But I think the bigger issue is the need to do this. Why not have a multi stage approach, instead of a single user input to do everything (the example given is “Can you send Bob the document he requested in our last meeting? Bob’s email and the document he asked for are in the meeting notes file”) you could have “get Bob’s email address from the meeting notes file” followed by “create a new email to that address” and “find the document” etc.

A major problem with many plans for ML systems is that they are based around automating relatively simple tasks. The example of sending an email based on meeting notes is a trivial task that’s done many times a day but for which expressing it verbally isn’t much faster than doing it the usual way. The usual way of doing such things (manually finding the email address from the meeting notes etc) can be accelerated without ML by having a “recent documents” access method that gets the notes, having the email address be a hot link to the email program (IE wordprocessor or note taking program being able to call the MUA), having a “put all data objects of type X into the clipboard (where X can be email address, URL, filename, or whatever), and maybe optimising the MUA UI. The problems that people are talking about solving via ML and treating everything as text to be arbitrarily parsed can in many cases by solved by having the programs dealing with the data know what they have and have support for calling system services accordingly.

The blog post suggests a problem of “user fatigue” from asking the user to confirm all actions, that is a real concern if the system is going to automate everything such that the user gives a verbal description of the problem and then says “yes” many times to confirm it. But if the user is at every step of the way pushing the process “take this email address” “attach this file” it won’t be a series of “yes” operations with a risk of saying “yes” once too often.

I think that one thing that should be investigated is better integration between services to allow working live on data. If in an online meeting someone says “I’ll work on task A please send me an email at the end of the meeting with all issues related to it” then you should be able to click on their email address in the meeting software to bring up the MUA to send a message and then just paste stuff in. The user could then not immediately send the message and clicking on the email address again would bring up the message in progress to allow adding to it (the behaviour of most MUAs of creating a new message for every click on a mailto:// URL is usually not what you desire). In this example you could of course use ALT-TAB or other methods to switch windows to the email, but imagine the situation of having 5 people in the meeting who are to be emailed about different things and that wouldn’t scale.

Another thing for the meeting example is that having a text chat for a video conference is a standard feature now and being able to directly message individuals is available in BBB and probably some other online meeting systems. It shouldn’t be hard to add a feature to BBB and similar programs to have each user receive an email at the end of the meeting with the contents of every DM chat they were involved in and have everyone in the meeting receive an emailed transcript of the public chat.

In conclusion I think that there are real issues with ML security and something like this technology is needed. But for most cases the best option is to just not have ML systems do such things. Also there is significant scope for improving the integration of various existing systems in a non-ML way.

• [1] https://simonwillison.net/2025/Apr/11/camel/