Angel Ceballos
SeraphicGuardian | Architect of Defensible Systems
Raleigh, North Carolina, United States
Actions
Angel Ceballos is an architect of defensible systems.
Her work focuses on how complex data and decision systems are designed, governed, and held accountable in real-world environments. Rather than treating governance as documentation or policy layered on top of technology, she builds it directly into system architecture.
With a foundation in data engineering and enterprise architecture, Angel designs and implements accountable infrastructure that governs risk, fairness, and long-term resilience. She challenges organizations to move beyond performance metrics and build systems that withstand scrutiny, degrade safely under stress, and produce defensible evidence by design.
Her mission is to ensure accountability and safety are not policy add-ons but engineered foundations that preserve the human voice, lived experience and narrative behind the number.
Links
Area of Expertise
Topics
Designing for Failure: Engineering AI Systems That Degrade Gracefully
AI systems operate in changing environments. Data shifts, assumptions evolve, dependencies update, and models continue running long after conditions have changed. When those changes are not anticipated in the system’s design, small issues can compound into larger consequences.
This session focuses on treating failure as a design consideration rather than an exception. We will examine how architectural decisions influence what happens when models drift, thresholds are misaligned, or upstream data changes unexpectedly. The discussion will cover containment strategies, escalation design, and approaches to limiting downstream impact when systems are under stress.
Participants will leave with a practical, systems-level perspective on building AI architectures that can absorb disruption, reduce unintended harm, and remain accountable when conditions change.
Abnormal Until Proven Otherwise
Security programs rely on labels that sound objective but are often not operationally defined: abnormal, high risk, and policy violation.
These labels appear in fraud monitoring, insider threat programs, incident response plans, access reviews, vendor risk workflows, AI-assisted classification, HR or legal escalation, privacy reviews, and compliance controls. They are not just descriptive terms. In many workflows, they determine what gets investigated, escalated, restricted, reported, or punished.
This talk argues that definition variance is an attack surface.
A user downloaded 400 files. That may be a fact. Calling the behavior abnormal is an interpretation. Calling the user high risk is a risk judgment. Calling the action a policy violation is an enforcement decision. Those are not the same thing.
When a security control depends on an undefined or inconsistently applied label, the organization may believe it has one control when it actually has multiple competing interpretations of that control. One analyst may treat a behavior as abnormal. Another may call it expected. One system may classify a user as high risk. Another may not. One team may treat an event as a policy violation while another sees the same event as authorized, explainable, or below threshold.
The issue is not only whether the event happened. It is whether the organization can defend the label applied to the event.
This session will walk through how the labels abnormal, high risk, and policy violation function as control points in security workflows: detection, prioritization, and enforcement. It will also include a short interactive exercise called The Label Lab, where attendees will work in small groups to review the same security scenario and decide which labels apply and what action should follow.
The exercise is designed to show how the same facts can produce different labels, different escalation paths, and different enforcement decisions when definitions are not operationalized.
Attendees will leave with a practical method for identifying undefined control labels, testing interpretation drift across teams, and strengthening security workflows before vague language becomes inconsistent enforcement, missed escalation, audit failure, or automated harm.
Depending on what the judges feel would be best, I am happy to present this material in either a 45-minute fashion or expand it to a half-day workshop. A half-day workshop would include an expanded version of the Label Lab Exercises and a hands-on Fix-It Lab to demonstrate what the Fix should be for when disagreements in labeling occur.
Detailed Outline
1. Opening thesis: labels are not facts
Security workflows often treat labels like abnormal, high risk, and policy violation as if they are objective facts.
They are not.
A security event usually contains at least two separate questions:
Did the event actually happen?
Does the event meet the organization’s definition of the label being applied?
Many workflows verify the first question and quietly assume the second.
Example: A user downloaded 400 files from an internal repository.
That does not automatically answer:
Was the behavior abnormal?
Was the user or event high risk?
Was the action a policy violation?
Those labels require definitions.
2. The three-label security chain
The talk focuses on three labels that commonly appear in security workflows:
Abnormal - detection
High risk - prioritization
Policy violation - enforcement
The argument is that these labels are not just language. They are control points.
Once a label triggers investigation, escalation, restriction, reporting, or punishment, the label becomes part of the control design.
If the definition varies, the control varies.
3. Where definition variance appears
Definition variance can appear in:
Fraud detection
Insider threat monitoring
Incident response escalation
Access reviews
Vendor risk scoring
Data classification
AI-assisted triage
HR, legal, or compliance referral workflows
Security awareness reporting
The same event can move through multiple teams or systems and receive different labels because each team, tool, vendor, or model applies a different interpretation.
4. Why this becomes a security risk
Definition variance can cause:
Missed escalations
Over-escalation
Inconsistent enforcement
Weak audit trails
Post-incident confusion
Vendor accountability gaps
Uneven or biased outcomes
AI systems applying undocumented standards
Controls that cannot be defended after the fact
The problem becomes more serious when vague labels are embedded into alerting logic, automated workflows, AI classification, access restrictions, compliance reporting, or disciplinary escalation.
5. Interactive exercise: The Label Lab
Attendees will work in small groups to review the same security scenario.
Scenario: The Late-Night Download
A user downloads 400 files from a shared internal repository at 11:48 p.m.
The files include: Customer records, Internal process documents, Pricing templates
Additional facts:
The user technically had access to the repository.
The user had never downloaded more than 30 files in a day before.
No DLP alert fired at the time.
Two weeks later, the user resigned.
Company policy prohibits “excessive downloading,” “unauthorized use,” and “activity that creates high risk to the organization.”
The policy does not define “excessive.”
A vendor tool labels the event “high risk,” but does not clearly explain the threshold.
Each group will answer: Label (and pick from Abnormal, High risk, or Policy violation); whether to Apply it (yes/no), Why; What action follows
Groups will also be asked:
What facts are confirmed?
What assumptions are being made?
What definition did you use?
Where would that definition be documented?
Who should own the definition?
What happens if another team disagrees?
The goal is to show that the facts may be stable while the labels and consequences vary.
Same facts. Different labels. Different controls.
6. Exercise debrief: when labels trigger consequences
The group exercise will be used to show how quickly disagreement appears.
Some groups may classify the event as abnormal but not a policy violation. Others may classify it as high risk because of the file volume, timing, and resignation. Others may hesitate because the user had access and the policy language is vague.
That disagreement is the point.
If different groups apply different labels to the same facts, and those labels trigger different actions, then the organization does not have one stable control. It has multiple competing interpretations of the control.
The debrief connects the exercise back to real workflows:
Alert triage
Insider threat escalation
Access review decisions
HR or legal referrals
AI-assisted risk scoring
Vendor security tools
Incident response documentation
7. How AI makes this worse
AI can scale undefined labels faster than humans can review them.
A model may classify something as abnormal, high risk, or a policy violation. But if the organization has not defined those labels clearly, the AI system may appear consistent while applying a standard no one can inspect, challenge, reproduce, or defend.
AI cannot reliably apply a standard the organization has not defined.
Security teams reviewing AI-assisted workflows need to ask whether the system is applying documented definitions or simply producing labels that look authoritative.
8. Practical controls
The session will close with practical ways to reduce definition variance:
Create operational definitions for high-impact labels
Separate factual verification from label interpretation
Define what evidence is required before a label can be applied
Version definitions like control logic
Make definitions accessible to humans and systems
Document who owns each definition
Test edge cases before automation
Require escalation paths when teams disagree
Audit outcomes for inconsistent application
Include definition checks in security reviews and AI governance reviews
9. Attendee takeaways:
A way to identify vague labels embedded in security workflows
A method for testing interpretation drift across teams
A checklist for separating event verification from classification
A framework for reviewing AI-assisted security decisions
A practical argument for treating definitions as part of security control design
Angel Ceballos
SeraphicGuardian | Architect of Defensible Systems
Raleigh, North Carolina, United States
Links
Actions
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top