Satheeshan Siva

📈 Tech, AI, magic, music, philosophy... one operating system, several outputs. 🇦🇺 Sydney.

[Satheeshan Siva @satheeshan]

There’s an industry murmur in AI security that says AI changes everything and your existing control library is obsolete.

I don’t think that’s true...

...Most of your controls still transfer.

The more interesting problem is that some controls can continue to pass even after AI has made the assumption underneath them false, e.g.

1. Change management assumes production changes originate inside your organisation.
2. Asset management assumes an asset has a stable, identifiable version.
3. Access recertification assumes every principal ultimately has a human owner.
4. Vendor management assumes the product you assessed is broadly the product you’ll still be using six months later.

AI breaks each of those assumptions without necessarily breaching the control itself.

So the dashboard stays green. The audit passes. The risk committee sees “effective”...

And the exposure is still there.

In the latest article in my AI security series, I look at where existing controls genuinely transfer, where they stop working, and the question I think every control owner should be asking:

"What did the original intent of the control assume that AI has made false?"

Tap the link below and read more:
https://t.co/SFaFvzsBwC

[Satheeshan Siva @satheeshan]

Most AI attack surface maps stop at the model or inference endpoint...

...the problem is that modern AI systems don’t.

Once you add tools, connectors, autonomous agents, credentials and human approval workflows, the attack surface expands... and some of the least mature controls sit above the point where most of these architecture diagrams end.

In the second article in this series, I’ve mapped the AI attack surface into nine layers:
Layer 1: Compute & hardware
Layer 2: Data pipeline
Layer 3: Training environment
Layer 4: Model artifacts
Layer 5: Deployment infrastructure
Layer 6: Inference runtime
Layer 7: Tool & connector layer
Layer 8: Agent layer
Layer 9: Human & process layer

For each layer, I help organisations to look at three questions:
1. What is it?
2. What does an attacker get from it?
3. And who owns it today?

That last question is usually the most revealing.

In my experience, layers 7 and 8 (tools, connectors and agents) are where ownership starts to become commonly difficult to pin down.

And once AI can take actions in production, the distinction between a security failure and a safety failure starts to matter a lot less operationally.

This article introduces the 9 layer reference map I’ll use for the rest of the series.

Tap the link below and read:
https://t.co/z2Oc5QQykV

[Satheeshan Siva @satheeshan]

If your AI security policy is an acceptable-use policy and a procurement questionnaire...

You don't actually have an AI security policy. You have part of one.

The problem is that “AI security” has become a catch-all term for four very different disciplines:

1. Securing AI systems
2. AI for security
3. AI safety
4. AI governance

And when you don't account for this, boards can walk away believing a risk has been addressed when nobody has actually asked the right question yet.

I unpack the distinctions between these disciplines in this first article, and introduce why assurance is the point where all four eventually meet.

Click the link and read the post for more...

https://t.co/RgScDvhPGV

[Satheeshan Siva @satheeshan]

Quite true...

https://t.co/kphX7HY3TV

[Satheeshan Siva @satheeshan]

Can't wait for this internet thing to come out... it looks pretty interesting...

https://t.co/teHcWsCVVU

[Satheeshan Siva @satheeshan]

@elonmusk Love the pure the integrity on display here

[Satheeshan Siva @satheeshan]

https://t.co/OQVZOOQRnV

This is amazing

[Satheeshan Siva @satheeshan]

Somehow 2024 was the year the whole world forgot that the whole point of lavalier mics is that you didn't have to hold them and/or you could easily hide them out of sight 😂

[Satheeshan Siva @satheeshan]

Crazy how much of the stuff sold on Amazon now is e-comm people reselling stuff from Temu and AliExpress.

[Satheeshan Siva @satheeshan]

@JoeBiden Or lower taxes for all and review government spending for efficiencies 🤷‍♂️

[Satheeshan Siva @satheeshan]

20/20 hindsight but the ongoing social, economic and health ramifications of excessive lockdown policies may likely be looked back as one of, if not the, largest global policy failures of our times.

https://t.co/5RQkte8yPR

[Satheeshan Siva @satheeshan]

Not sure how I’m feeling about Meta AI as the default in IG search

[Satheeshan Siva @satheeshan]

One of the most powerful skills to have is pattern recognition

[Satheeshan Siva @satheeshan]

The paradox of choice: the more options we have, the less likely we are to decide on any one option

[Satheeshan Siva @satheeshan]

Process > outcome

[Satheeshan Siva @satheeshan]

Test, learn, optimise, rinse and repeat

[Satheeshan Siva @satheeshan]

The gift of social media is that it has allowed us to connect in a way like never before possible.

But it could be argued that its tradeoff is that it has led to us over-optimize on appearance, stage presence and smoke and mirrors… and under-optimize on character, value systems and behaviours.

The latter being those things that really speak to who we are in the long run.

[Satheeshan Siva @satheeshan]

@Support option for premium signup doesn't show on my account

[Satheeshan Siva @satheeshan]

@premium @premium I don't seem to have the option to signup for Premium

[Satheeshan Siva @satheeshan]

So I was thinking of putting that new Jacuzzi in just about… here… @ Ta Prohm https://t.co/Ni0xxj6ahI