For two decades, security practitioners have lived with a tidy, almost childlike definition of an incident.
“An incident is any unauthorized access, use, disclosure, modification, or destruction of information.”
It shows up in frameworks, incident response (IR) handbooks, compliance guidelines, and every tabletop exercise your CISO forces you into:
This worked fine when data lived in one place, was governed by one company, and was inside one jurisdiction. Investigations were straightforward. Access meant access. Logs were logs. A breach was a breach.
That world is gone.
Between regional cloud architectures, data locality laws, sovereign cloud commitments, and AI systems that process data in places you can’t pronounce, the old definitions no longer describe reality. And when I recently asked a group of analysts a very simple question, “Does sovereign data count as part of an incident if we’re not legally allowed to inspect it?” The room froze.
Not because the question was shocking. Because the honest answer is: our current definitions don’t work anymore.
But why?
Legacy definitions: clear, simple, and wrong
Traditional definitions of an incident assume three things:
- You know where your data lives.
- You’re legally allowed to access it.
- Your investigation can involve extracting, moving, or exporting it.
These assumptions fall apart the moment your data enters:
- A sovereign cloud region that prohibits cross-border access.
- A regulated environment where metadata itself is restricted.
- An AI system where processing and storage locations don’t match.
- A multi-tenant service where the vendor controls the infrastructure but not the laws that govern it.
The biggest surprise is not that the definitions are outdated. It’s that we’re still pretending they aren’t.
Sovereign data creates a new category of incident
Here’s the paradox sovereignty introduces: You must investigate the thing you’re legally forbidden from touching.
Your IR plan probably includes:
- Pulling logs
- Reviewing access patterns
- Exporting evidence to your SIEM
- Reconstructing what happened
But sovereignty says:
- You cannot export the logs
- Some logs can’t exist
- Evidence cannot cross a border
- Access is restricted even from your own IR team
So let’s ask the previously unthinkable question: If a security event involves sovereign data, but that data cannot be accessed or moved to determine scope, does it still meet the definition of an incident?
Regulators say yes. Sovereignty rules often say no. Your CISO says, “Ask legal.” Legal says “ask the cloud provider.” The cloud provider says, “Check the documentation.” The documentation says, “It depends.”
This is where modern IR falls apart.
AI makes this incredibly messy
Before sovereignty challenged incident definitions, AI finished the job.
AI systems routinely break the neat boundaries our definitions rely on. A prompt occurs in region A, while the processing happens in region B. The embeddings live in region C, the model weights are globally distributed, and the vendor’s FAQ says “we don’t store your data except for when we do.” And then that AI generates new data, reasoning, and outputs – depending on where it’s based, that might shift regions yet again.
Even when providers claim no training on customer inputs, the metadata, embeddings, and logs often remain, sometimes in regions subject to laws different from those of the data origin.
So what happens when EU personal data is sent to an AI service running in the US, a regulated prompt is processed in a region the customer didn’t choose, or an embedding is cached outside the sovereignty boundary?
Is it considered an incident, policy violation, sovereignty breach, or simply “expected model behavior”?
Under legacy definitions, we genuinely don’t know.
It’s time for new categories, not new wordsmithing
Security can’t fix this with better phrasing. Legal can’t solve it with more disclaimers. Cloud vendors can’t solve it by drawing regional maps in marketing decks. We need a new taxonomy of incidents, one that reflects the world we actually work in.
1. Operational security incidents
Traditional unauthorized access, misuse, or compromise.
2. Sovereignty violations
Any action, benign or malicious, that touches data in a region where the investigator is restricted by law or contractual boundary.
These are not data breaches as traditionally defined, but they are absolutely material events.
3. AI processing violations
Cross-region inference, embedding leakage, prompt exposure, or processing outside declared boundaries.
This covers everything from model misuse to unexpected data persistence in AI infrastructure.
4. Silent processor incidents
My name for the category nobody wants to talk about:
Something happened to data that you cannot investigate because you are legally barred from looking.
This isn’t fiction, it’s the lived reality of modern regulated environments.
5. Multi-jurisdiction metadata incidents
Events where the metadata violates boundary rules, even when the data does not.
This will be one of the biggest regulatory fights of the next decade.
What this means for detection and response
Organizations need to update their IR programs to include:
- Region-specific IR workflows
- Constraints-based forensics (what you can’t touch matters as much as what you can)
- Separate SIEM/analytics stacks for sovereign environments
- A formal AI incident classification matrix
- Legal/regulatory decision trees connected to IR – not adjacent to it
- Vendor transparency requirements for inference, caching, and embeddings
And yes, vendors have to evolve too. Vendors must provide:
- Region-locked processing
- Region-locked storage
- Clear boundaries around AI components
- Documented sub-processor locations
- Transparent guarantees about training, caching, logs, and embeddings
- IR tooling inside the region where customers can access it
Without this, customers cannot meet sovereignty requirements and cannot perform real incident response.
Where we go from here
My original question wasn’t a trap. It was a test of whether the industry is ready to confront a truth we’ve avoided:
Our definition of an incident was written for a world that no longer exists.
Today’s data landscape, sovereignty, AI, regional clouds, and cross-border inference demands a more nuanced, jurisdiction-aware, architecture-aware model of incident handling. Security must now focus on unauthorized context, location, processing, and boundaries.
The future of IR is intersectional: part security, part compliance, part data governance, part AI architecture, part geopolitical reality. We either evolve our definitions now, or we’ll keep having rooms full of analysts staring back at us like we just asked them to define quantum physics.
And frankly, the attackers aren’t waiting for us to catch up.
