Breaking the Cycle: From Red Teaming to DevSecOps Leadership

Speaker: 00:06
Absolutely. It’s becoming fundamental.

Speaker 1: 00:07
But you know, with all that cool stuff, there’s this huge question, maybe when we don’t ask enough, how do we actually keep it safe, secure?

Speaker: 00:16
Precisely. And as these AI systems get, well, smarter and more embedded, the ways things can go wrong, they just multiply.

Speaker 1: 00:24
Accidentally or worse, deliberately.

Speaker: 00:26
Yeah. The risks feel different, maybe more complex than the usual software security we’re used to.

Speaker 1: 00:31
And that’s exactly what we’re digging into today. We’ve looked at sources on AI security frameworks specifically for enterprises.

Speaker: 00:36
Aaron Powell Think of them like um blueprints or maybe instruction manuals for AI safety.

Speaker 1: 00:42
Aaron Powell Yeah, good analogy. They’re not just checklists, though. They’re structured ways to find the weak spots, set rules, deal with laws, and importantly, build trust.

Speaker: 00:51
Aaron Ross Powell Right. Our mission here is to pull out the key insights from these sources. We’ll be talking about NIST, Microsoft’s Take, a MITREATLES, and uh Databricks Das F.

Speaker 1: 01:01
Different angles, but all aiming for the same thing, safeguarding AI.

Speaker: 01:05
Aaron Ross Powell So NIST is often the first one people mention, the AI risk management framework.

Speaker 1: 01:10
Right. It’s meant to be pretty comprehensive, covering the whole AI lifespan.

Speaker: 01:14
Yeah. And what’s really core to NIST is this cyclical idea. It’s got four main functions. First, you map out your AI systems, figure out where the risks might pop up.

Speaker 1: 01:23
Okay, Matt, got it.

Speaker: 01:24
Then you measure those risks you found. How likely? How bad could it be?

Speaker 1: 01:29
Makes sense.

Speaker: 01:30
Next is manage. That’s putting controls in place to deal with the risks. And finally, maybe most crucially, you govern the whole thing.

Speaker 1: 01:37
Govern, like setting up teens, policies.

Speaker: 01:40
Exactly. Oversight, making sure it’s continuous. It’s less about a single check and more about weaving risk management into everything from the start through deployment and beyond.

Speaker 1: 01:50
And it really pushes for that team approach. Security, le legal, devs, business folks, all involved early.

Speaker: 01:56
Absolutely essential.

Speaker 1: 01:57
Okay, what about Microsoft? Their framework. Obviously, they have their own ecosystem, but the principles seem broader.

Speaker: 02:02
Aaron Powell They do. And uh a key thing for Microsoft is that they look beyond just the technical security holes.

Speaker 1: 02:08
Right. I saw that. They talk security, yes, but also privacy, fairness, transparency, accountability.

Speaker: 02:14
Exactly. It’s about operationalizing all of those together. Protecting against attacks is basic, sure. But things like is the AI fair? Can we understand its decisions? Who’s responsible if it messes up?

Speaker: 02:30
They do. It’s a more holistic view of secure.

Speaker 1: 02:32
But then there are the specific attacks on AI. That seems like INRE atlas territory.

Speaker: 02:36
Yeah, ATLS is all about the adversarial side, understanding how attackers might specifically target AI systems.

Speaker 1: 02:44
Right. So how do you secure something designed to change?

Speaker: 02:46
That’s the challenge. AI vulnerabilities are different. ATLS actually catalogs attack types, like um data poisoning.

Speaker 1: 02:53
That’s not just bad code, is it?

Speaker: 02:55
No, it’s messing with the training data itself, teaching the AI the wrong things, subtly. Or evasion attacks, creating inputs like slightly tweaked images that fool the AI but look fine to us.

Speaker 1: 03:06
Wow. So ATLS helps you think like an attacker.

Speaker: 03:09
Pretty much. It’s a playbook to test your defenses against these unique AI threats, build better walls by knowing how they might be breached.

Speaker 1: 03:17
Okay, and the last one we looked at was Databricks, the DDSF.

Speaker: 03:20
Right, the Databricks AI security framework. This one often builds on NIST and MITRE, mapping concepts across. It gives you a catalog of risks and controls quite specific to common machine learning workflows.

Speaker 1: 03:33
So more hands-on, maybe.

Speaker: 03:34
You could say that. It really focuses on actionable controls for specific risks. Things like data protection, keeping training data safe, model security protecting the model itself, access control, monitoring, uh, the practical stuff.

Speaker 1: 03:47
Aaron Powell So if you’re running big data science platforms, DSF helps translate the theory into actual steps.

Speaker: 03:52
That seems to be the goal, yeah. Making it practical for implementation.

Speaker 1: 03:55
Aaron Powell So none of these exist on their own, really. You mentioned banking earlier. How might a bank use these?

Speaker: 04:01
Good example. Banking is high stakes, credit, fraud, fairness regulations. They could use NIST for the overall structure, right? Yeah. AI uses in loans, measure bias risks or evasion threats, manage with audits and testing, govern with cross-functional teams.

Speaker 1: 04:17
Aaron Powell Okay. NIST provides the skeleton.

Speaker: 04:19
Exactly. Then they might pull in Microsoft’s principles to really focus on fairness in those loan algorithms.

Speaker 1: 04:25
And use MITRE Atlas to understand how fraudsters might try to trick their AI detection systems.

Speaker: 04:31
Precisely. Layering them together makes sense.

Speaker 1: 04:33
Aaron Powell So the big takeaway isn’t pick one, is it?

Speaker: 04:36
Aaron Powell Not usually. No. Most places seem to mix and match, tailor it to their own situation, their industry, the tech they’re using.

Speaker 1: 04:44
Aaron Powell And it’s definitely not a one-and-done project.

Speaker: 04:46
Oh, absolutely not. It’s a continuous journey. Needs that coordinated effort, security, devs, legal management, plus constant checking, updating, adapting as the AI and the threats evolve.

Speaker 1: 04:57
Aaron Powell So for you listening, these frameworks are really essential guides, blueprints, like we said, for managing AI risks through the whole process.

Speaker: 05:05
They help build safer AI, meet compliance, protect data, and maybe most importantly, build that crucial trust.

Speaker 1: 05:12
And maybe a final thought to leave you with As AI gets deeper into critical areas like finance or healthcare, what does it really mean to build AI that’s not just secure from attack, but truly trustworthy? You know, upholding fairness, being transparent, even when things get tough or someone’s actively trying to break it?

Speaker: 05:30
That’s the deeper challenge, isn’t it? Building trustworthy AI.

Practical DevSecOps (a training division of Hysn Technologies Inc) provides world-class, practical, and hands-on Product Security training and certification programs. Our state-of-the-art online lab ensures our students learn the practical aspects of the course and showcase their knowledge to employers and colleagues with world-renowned Certifications.