Audience pulse
Where does AI help you today?
Select the closest answer. There is no wrong starting point.
01Start with a real task
02Package what repeats
03Add agency only when earned
Session 1 · Foundations
From chat.
To repeatable capability.
How LLMs, projects, skills, tools and agents change the way cybersecurity work gets delivered.
60 minutesNo engineering prerequisiteInteractive
Repeated
security task
security task
Prompt
Skill
Agent
Evidence
Deliverable
space / click to reveal
Mental model
An LLM predicts a useful continuation.
It transforms input into tokens, uses learned patterns plus current context, and generates the next token repeatedly.
OpenAI API conceptsChoose the next token
The suspicious sign-in should be _____
Illustrative probabilities. Context changes the distribution.
Prompt anatomy
The model sees tokens.
The application must preserve trust.
System
Operating boundariesRole, output contract, prohibited actions and escalation rules.
User
Requested outcomeWhat the user wants accomplished in this interaction.
Context
Evidence and historyDocuments, retrieved passages, prior decisions and run state.
Tools
Observed resultsStructured results returned by approved capabilities.
Untrusted data
Never implicit authorityEmail, web pages and uploads may contain hostile instructions.
The first upgrade
Chat solves a moment.
A project preserves working context.
Individual chat
Cold start
- Re-explain scope and standards
- Manually attach the right files
- Copy results into the next task
- Quality depends on the operator
Project / workspace
Persistent context
- Shared instructions and reference files
- Conversation continuity
- Less repeated setup
- Still human-directed, step by step
A workspace remembers the work. It does not yet govern the work.
The maturity ladder
Add structure before autonomy.
01
Chat
One conversation
02
Project
Persistent workspace
03
Skill
Packaged expertise
04
Tool
Callable action
05
Agent
Bounded loop
06
Workflow
Governed execution
07
Multi-agent
Specialist system
Best for exploration, explanation and one-off analysis. Human provides context and owns every transition.
Candidate scorecard
A good first candidate is bounded, repeated and reviewable.
Candidate
Repeats
Clear output
Easy to review
Safe scope
Architecture review draft
High
Findings schema
Architect validates
Read-only
KQL query drafting
High
Query + metadata
Needs dry-run
Dev workspace
Deploy controls to production
Variable
Many side effects
Hard to reverse
High impact
Start: architecture reviewStart: document packStart: schema validationLater: production writes
The reusable unit
A skill packages the task—not merely the prompt.
A versioned bundle can carry instructions, specialist references, output schemas, deterministic scripts and reusable assets.
OpenAI Skills documentationsecurity-architect/
├── SKILL.md # trigger + workflow
├── references/
│ ├── threat-modeling.md
│ └── cloud-security.md
├── schemas/
│ └── findings.schema.json
├── scripts/
│ ├── iac_scan.py
│ └── report_html.py
└── assets/
└── report-template/
├── SKILL.md # trigger + workflow
├── references/
│ ├── threat-modeling.md
│ └── cloud-security.md
├── schemas/
│ └── findings.schema.json
├── scripts/
│ ├── iac_scan.py
│ └── report_html.py
└── assets/
└── report-template/
Interactive X-ray
Inside a production cybersecurity skill.
Selected component
SKILL.md · the operating contract
Defines when the skill applies, its workflow, boundaries, evidence rules and definition of done.
It routes work. It does not contain every reference inline.
Hybrid execution
Expert judgment and deterministic code do different jobs.
Agent judgment
Interpret context
- Identify trust boundaries
- Reason about abuse paths
- Explain business impact
- Prioritize remediation
Deterministic scripts
Guarantee the floor
- Validate JSON Schema
- Scan known IaC patterns
- Calculate and transform
- Render stable documents
Assessmodel
Validatecode
Reviewhuman / policy
Rendercode
Capability map
Tools, MCP and retrieval solve different problems.
Model
Requests a named action
Arguments follow a schema.
→
Application
validate_findings()
Code authenticates, authorizes, executes and returns a structured result.
→
Result
Pass / fail
The model receives data—not implicit authority.
Agent host
Discovers approved capabilities.
→
Model Context Protocol
Standard messages for tools, resources and prompts.
→
Approved servers
MS Learn · ArchStudio · internal knowledge
Question
What evidence is relevant?
→
Retriever
Searches an indexed corpus.
→
Grounded context
Relevant passages + citations reach the model.
PLANnext step
ACTrequest tool
OBSERVEread result
CHECKdone / retry
Bounded
agent
agent
The loop
Reason. Act. Observe.
Stop safely.
Bounds
Tool allowlist · iteration cap · token budget · timeout
Contracts
Structured input · structured tool arguments · validated output
Escalation
Needs input · needs review · blocked · failed safely
Do not anthropomorphize
The difference is controlled action—not personality.
Property
Chat assistant
Agent
Governed workflow
Primary behavior
Respond
Choose and act
Execute controlled stages
State
Conversation
Working memory
Durable run state
Tools
Optional
Core capability
Scoped per node
Control
Human drives turns
Runtime limits
Gates, policy, evidence
Agency starts when the system may choose an action. Governance starts when that choice is controlled and recorded.
Multi-agent patterns
Choose who owns the final answer.
Managerowns response
Architectagent as tool
Complianceagent as tool
Detectionagent as tool
Best when one component must combine specialist outputs and enforce shared controls.
Triageroutes
Specialisttakes control
Direct replyspecialist owns turn
Best when a specialist should directly own the interaction after routing.
Codeselects route
Specialistbounded task
Gatedeterministic
Best when sequence, approvals and failure behavior must be explicit and testable.
Complexity gate
More agents do not automatically create more value.
Stay single-agent when
- One prompt and a few tools solve the task
- All steps use the same context
- There is one definition of done
- Specialists would only repeat each other
Consider multi-agent when
- Expert roles have different evidence and instructions
- Work can proceed independently
- Hand-offs are typed artifacts
- An independent reviewer adds real assurance
+Specialization
+Parallel work
−Coordination cost
−Larger attack surface
Decision practice
Match the architecture to the task.
Task
Start with
Why
Upgrade trigger
Explain one Sentinel alert
Chat / project
Human owns context
Repeated triage procedure
Architecture review report
Skill + tools
Stable method and output
Multi-step evidence collection
Greenfield delivery pack
Governed workflow
Parallel specialists + artifacts
Independent assurance
Deploy controls automatically
Later-stage agent
High side effects
Policy, simulation, approval, rollback
A practical 90-day path
Adopt capability in measured increments.
Week 1
Select one repeated, bounded task.
Week 2
Baseline quality, time and rework.
Week 3–4
Package a reviewed skill and schema.
Month 2
Add read-only tools and deterministic checks.
Month 3
Add gates, traces and adversarial tests.
Decision
Scale only when evidence supports it.
Measure qualityMeasure review effortMeasure defects caughtMeasure safe failure
Session 1 synthesis
Package expertise.
Bound action. Measure value.
01 · Start
Repeated task
Not a fascination with agents.
02 · Build
Skill before autonomy
Make expertise versioned and reviewable.
03 · Scale
Governed workflow
Only when controls and evidence grow with capability.
Next: what changes when AI can read untrusted data and take action?