Schneier on Security

Clive Robinson • September 1, 2026 5:37 PM

@ Bruce, ALL,

The first question to answer is,

1, Does there have to be a reason?

And if you can think of one the next question is,

2, Does it effect me?
3, How?
4, What actions to take.

The problem with the first question is it’s a “random rabbit hole” with a near infinite warren behind it.

Thus you need to find a way to enumerate things.

The problem, if it were a human behind it there would be a number of tricks you could use. But an AI is unlikely to respond in a way that would let you progress.

Whilst I did not have to deal with AI I realised I did not have the time or the will to play Email Games.

So as you might remember I cut through 2&3 and as a solution to 4 I stopped using personal email all together.

Eventually other people will realise this is the only sensible option with AI now in the email spamming game.

It’s not hard to see how AI will kill off not just email in fairly short order but nearly all unauthenticated electronic communications that does not have a significant cost to the Agentic herders.

If people think “Nuisance phone calls” are bad wait until AI does a number on all your electronic communications, it will make DDoS look like a pleasantry…

It’s why after “job agents” tried tracking me down with offers they thought I could not refuse I dumped all personal Email into an auto-responder for a while and then pulled the plug.

Clive Robinson • September 2, 2026 11:53 AM

@ Bruce, ALL,

The post #comment-457422 from “Clive” was not as far as I’m aware from,

“The me, me :-)”

But hey I’ve many doubles[1] as has been seen in the past.

But getting back to your point of,

“but someone automated the process.”

True enough, but does that have anything to do with you specifically? And was it even deliberate?

Let me put it this way tools get created and made, but how they get used is not upto the designer/creator of the tool but the using entity.

Thus people have been known to use screwdrivers as hammers or even tooth picks…

But I’m reminded of something from my past over half a century ago.

Back then Fizzy drinks came in glass bottles that had “refunds” on them and had special wooden crates for shops to return the empties back to the bottling plant.

Now as many kids of that time knew if you blew across the top of such a glass bottle at the right angle you would get a tone from it as the circulating air in the bottle created a resonance. The bottle was not designed for that but it was a tool like any other wind instrument.

My local “corner shop” –that was not actually on a corner– had an alley way between it and another shop in the row. At certain times the wind would blow forcefully down it and cause the wind to blow across the wooden crates of return bottles.

And yes an eerie noise would arise and be very noticeable at night thus giving rise to people saying the place was haunted.

The point is the bottle was in effect a wind instrument thus a “tool” but it was not created as such. The fact that crates of them got stacked in the shops yard had noting what so ever to do with them being “tools” or getting used as such when the wind happened to blow from a given direction.

Now we could argue that the AI has “purpose” and this has taken an existing “tool” and used it.

But has it?

Agentic AI is a stochastic process using another stochastic process.

It’s also known that directing Agentic AI “safely” is not just difficult it’s thought by some to actually be not possible[2]. And it was something I was working on producing a proof for[3], before my recent medical issues putting a gimp in things.

[1] Though it would not be the first or last time another “Clive Robinson” from the UK has shown up. A couple of months back when I showed how a lawyer looking for an expert witness could get misled by AI where the AI confused me with another Clive Robinson through a “Companies House” listing was mildly amusing but made a point.

[2] Put overly simply it is the “observer problem” again. We now sort of except the proof that,

“Guardrails will always fail to chosen prompt attacks”

Some for fiscal reasons are fighting a regard action but,

“Their goose is well and truly cooked”.

[3] I’ve outlined why this is so previously mentioning the works of Claude Shannon and Gus Simmons. It’s for the simplest reason “redundancy is required to communicate” and “where there is observed redundancy there is uncertainty”. Thus even simple crypto can hide things from the Guardrail “observer” but leave it fully understandable by the LLM system.

We also know that the control systems of agentic AI systems are in reality just another “guardrail” system no better, no worse, it is just a very imperfect observer with a go/no-go switch. So the reality is, it’s a turtles all the way down issue.