Internal · shadow AI levels redesign

Six reworked + three new directions

The first six (V1, V2, V3, V5, V6, V7) replace the earlier Claude-y versions and now follow the de-claude-ified grammar from the open graphics-improvements PR: edge-to-edge rows separated by thin slate-200 rules, mono category codes in the gutter, no pastel pills, no rounded floaty cards. V8, V9, V10 are net-new.

V1 V2 V3 V5 V6 V7 V8 V9 V10

Reworked

V1 · Editorial radar

Single axis, real tick marks, no rounded gradient bars. Plotted nodes carry their level code; bodies live in an editorial flat list below.

Blast radius · lowCritical

LV01 · Naive

LV02 · Convenience

LV03 · Defiant

LV04 · Embedded

LV05 · Agentic

LV06 · Supply Chain

LV01LOW 1/5
Naive
Employees use ChatGPT, Claude, and Gemini because no one told them not to. Roughly 25 to 30 percent of all shadow AI usage. Data leakage risk stays substantial even when intent is harmless.
Modulos directDetection: SaaS discovery
LV02MED 2/5
Convenience
Enterprise Copilot exists, but staff default to consumer ChatGPT because it is faster. 30 to 35 percent of usage. Sensitive data leaves the corporate perimeter under different governance.
Modulos orchestratesDetection: DLP / data flow
LV03MED 3/5
Defiant
Senior staff who know the policy and bypass it anyway. Leadership often the heaviest user. Training stops working. Only infrastructure enforcement plus regulatory urgency changes behaviour.
Modulos orchestratesDetection: SSE / CASB
LV04HIGH 4/5
Embedded
Notion AI, Copilot in Excel, Zoom summaries, Grammarly were all approved before they added AI. The AI itself was never assessed. By 2026 ~70% of enterprise AI happens inside previously-approved SaaS.
Modulos orchestratesDetection: Embedded-AI SPM
LV05CRIT 5/5
Agentic
Autonomous agents act on systems, data, and decisions at machine speed. 47 percent of enterprises report an agent security incident in the past year. CVE-2025-53773 showed how prompt injection could enable full system compromise.
Modulos with partnersDetection: Agent runtime trust
LV06HIGH 4/5
Supply Chain
Your legal vendor summarises contracts with AI. Your CRM added AI insights. EU AI Act value-chain provisions, DORA, NIS2, and customer commitments turn these into governable risk regardless.
Modulos directDetection: TPRM

Reworked

V2 · Severity ladder

Vertical thin meter on the left rule fills with severity. Right column reads like a journal byline: detection, Modulos code, prose.

LV01LOW 1/5
Naive
“I didn't know there was a policy.”
Employees use ChatGPT, Claude, and Gemini because no one told them not to. Roughly 25 to 30 percent of all shadow AI usage. Data leakage risk stays substantial even when intent is harmless.
DetectionSaaS discovery
ModulosDIR
EU AI Act Article 4 literacy training, policy management, intake form for declared use cases, and Scout discovery across GitHub, cloud, and Atlassian.
LV02MED 2/5
Convenience
“The approved path is too slow.”
Enterprise Copilot exists, but staff default to consumer ChatGPT because it is faster. 30 to 35 percent of usage. Sensitive data leaves the corporate perimeter under different governance.
DetectionDLP / data flow
ModulosORC
Quantifies the monetary risk of unapproved usage. Intake plus Scout analysis makes the approved path fast enough to compete with shadow usage.
LV03MED 3/5
Defiant
“The policy is wrong.”
Senior staff who know the policy and bypass it anyway. Leadership often the heaviest user. Training stops working. Only infrastructure enforcement plus regulatory urgency changes behaviour.
DetectionSSE / CASB
ModulosORC
Regulatory framing compliance officers need to escalate (EU AI Act fines, board liability) and an audit trail of override-with-justification events.
LV04HIGH 4/5
Embedded
“The tool was approved. The AI inside it wasn't.”
Notion AI, Copilot in Excel, Zoom summaries, Grammarly were all approved before they added AI. The AI itself was never assessed. By 2026 ~70% of enterprise AI happens inside previously-approved SaaS.
DetectionEmbedded-AI SPM
ModulosORC
Embedded-AI signals feed into intake. Scout risk-classifies each newly-surfaced capability and re-maps controls without waiting for procurement.
LV05CRIT 5/5
Agentic
“The agent was approved. Its actions weren't.”
Autonomous agents act on systems, data, and decisions at machine speed. 47 percent of enterprises report an agent security incident in the past year. CVE-2025-53773 showed how prompt injection could enable full system compromise.
DetectionAgent runtime trust
ModulosPRT
Agent trust scores ingest as risk evidence. Modulos classifies agents under EU AI Act high-risk categories and continuously monitors them.
LV06HIGH 4/5
Supply Chain
“Our vendor is using AI on our data.”
Your legal vendor summarises contracts with AI. Your CRM added AI insights. EU AI Act value-chain provisions, DORA, NIS2, and customer commitments turn these into governable risk regardless.
DetectionTPRM
ModulosDIR
Third-party AI risk handled inside the same control framework. A single control satisfies obligations across the EU AI Act, ISO 42001, NIST AI RMF, NIS2, and DORA in parallel.

Reworked

V3 · Reality / Modulos columns

No dark/light card split. Three editorial columns per row separated by thin rules: level meta, the reality, the Modulos response.

LV01LOW 1/5
Naive
“I didn't know there was a policy.”
The reality
Employees use ChatGPT, Claude, and Gemini because no one told them not to. Roughly 25 to 30 percent of all shadow AI usage. Data leakage risk stays substantial even when intent is harmless.
Detection layer · Shadow-app surfacing
Modulos · DIR
EU AI Act Article 4 literacy training, policy management, intake form for declared use cases, and Scout discovery across GitHub, cloud, and Atlassian.
Regulatory · EU AI Act Art. 4
LV02MED 2/5
Convenience
“The approved path is too slow.”
The reality
Enterprise Copilot exists, but staff default to consumer ChatGPT because it is faster. 30 to 35 percent of usage. Sensitive data leaves the corporate perimeter under different governance.
Detection layer · DLP / data flow
Modulos · ORC
Quantifies the monetary risk of unapproved usage. Intake plus Scout analysis makes the approved path fast enough to compete with shadow usage.
Regulatory · EU AI Act Art. 4 · GDPR
LV03MED 3/5
Defiant
“The policy is wrong.”
The reality
Senior staff who know the policy and bypass it anyway. Leadership often the heaviest user. Training stops working. Only infrastructure enforcement plus regulatory urgency changes behaviour.
Detection layer · SSE / CASB
Modulos · ORC
Regulatory framing compliance officers need to escalate (EU AI Act fines, board liability) and an audit trail of override-with-justification events.
Regulatory · EU AI Act fines · Board liability
LV04HIGH 4/5
Embedded
“The tool was approved. The AI inside it wasn't.”
The reality
Notion AI, Copilot in Excel, Zoom summaries, Grammarly were all approved before they added AI. The AI itself was never assessed. By 2026 ~70% of enterprise AI happens inside previously-approved SaaS.
Detection layer · Embedded-AI SPM
Modulos · ORC
Embedded-AI signals feed into intake. Scout risk-classifies each newly-surfaced capability and re-maps controls without waiting for procurement.
Regulatory · EU AI Act Art. 50 · Annex III
LV05CRIT 5/5
Agentic
“The agent was approved. Its actions weren't.”
The reality
Autonomous agents act on systems, data, and decisions at machine speed. 47 percent of enterprises report an agent security incident in the past year. CVE-2025-53773 showed how prompt injection could enable full system compromise.
Detection layer · Agent runtime trust
Modulos · PRT
Agent trust scores ingest as risk evidence. Modulos classifies agents under EU AI Act high-risk categories and continuously monitors them.
Regulatory · EU AI Act high-risk · NIS2
LV06HIGH 4/5
Supply Chain
“Our vendor is using AI on our data.”
The reality
Your legal vendor summarises contracts with AI. Your CRM added AI insights. EU AI Act value-chain provisions, DORA, NIS2, and customer commitments turn these into governable risk regardless.
Detection layer · TPRM
Modulos · DIR
Third-party AI risk handled inside the same control framework. A single control satisfies obligations across the EU AI Act, ISO 42001, NIST AI RMF, NIS2, and DORA in parallel.
Regulatory · EU AI Act value chain · DORA · NIS2

Reworked

V5 · Dense data sheets

V3-C dense card grammar without the pastel pills. Each cell shows the hero metric, the level's short quote, body, and a small definition list of detection + Modulos role.

LV01LOW 1/5

Naive

“I didn't know there was a policy.”

25–30%of shadow AI usage

Employees use ChatGPT, Claude, and Gemini because no one told them not to. Roughly 25 to 30 percent of all shadow AI usage. Data leakage risk stays substantial even when intent is harmless.

Detection: SaaS discovery
Modulos: DIR · direct

LV02MED 2/5

Convenience

“The approved path is too slow.”

30–35%of shadow AI usage

Enterprise Copilot exists, but staff default to consumer ChatGPT because it is faster. 30 to 35 percent of usage. Sensitive data leaves the corporate perimeter under different governance.

Detection: DLP / data flow
Modulos: ORC · orchestrates

LV03MED 3/5

Defiant

“The policy is wrong.”

5–10%leadership-heavy

Senior staff who know the policy and bypass it anyway. Leadership often the heaviest user. Training stops working. Only infrastructure enforcement plus regulatory urgency changes behaviour.

Detection: SSE / CASB
Modulos: ORC · orchestrates

LV04HIGH 4/5

Embedded

“The tool was approved. The AI inside it wasn't.”

~70%of AI interactions

Notion AI, Copilot in Excel, Zoom summaries, Grammarly were all approved before they added AI. The AI itself was never assessed. By 2026 ~70% of enterprise AI happens inside previously-approved SaaS.

Detection: Embedded-AI SPM
Modulos: ORC · orchestrates

LV05CRIT 5/5

Agentic

“The agent was approved. Its actions weren't.”

47%had incidents

Autonomous agents act on systems, data, and decisions at machine speed. 47 percent of enterprises report an agent security incident in the past year. CVE-2025-53773 showed how prompt injection could enable full system compromise.

Detection: Agent runtime trust
Modulos: PRT · with partners

LV06HIGH 4/5

Supply Chain

“Our vendor is using AI on our data.”

Everymodern vendor

Your legal vendor summarises contracts with AI. Your CRM added AI insights. EU AI Act value-chain provisions, DORA, NIS2, and customer commitments turn these into governable risk regardless.

Detection: TPRM
Modulos: DIR · direct

Reworked

V6 · Compliance worksheet

Edge-to-edge worksheet, thin slate-200 rules between rows, tabular-nums everywhere. Designed to be scanned by a compliance officer, not a marketing buyer.

Level

What it looks like

Detection

Risk

Modulos

Reg hit

LV01LOW 1/5

Naive

Employees use ChatGPT, Claude, and Gemini because no one told them not to. Roughly 25 to 30 percent of all shadow AI usage. Data leakage risk stays substantial even when intent is harmless.

Shadow-app surfacing

1/5

DIR

EU AI Act Art. 4

LV02MED 2/5

Convenience

Enterprise Copilot exists, but staff default to consumer ChatGPT because it is faster. 30 to 35 percent of usage. Sensitive data leaves the corporate perimeter under different governance.

DLP / data flow

2/5

ORC

EU AI Act Art. 4 · GDPR

LV03MED 3/5

Defiant

Senior staff who know the policy and bypass it anyway. Leadership often the heaviest user. Training stops working. Only infrastructure enforcement plus regulatory urgency changes behaviour.

SSE / CASB

3/5

ORC

EU AI Act fines · Board liability

LV04HIGH 4/5

Embedded

Notion AI, Copilot in Excel, Zoom summaries, Grammarly were all approved before they added AI. The AI itself was never assessed. By 2026 ~70% of enterprise AI happens inside previously-approved SaaS.

Embedded-AI SPM

4/5

ORC

EU AI Act Art. 50 · Annex III

LV05CRIT 5/5

Agentic

Agent runtime trust

5/5

PRT

EU AI Act high-risk · NIS2

LV06HIGH 4/5

Supply Chain

Your legal vendor summarises contracts with AI. Your CRM added AI insights. EU AI Act value-chain provisions, DORA, NIS2, and customer commitments turn these into governable risk regardless.

TPRM

4/5

DIR

EU AI Act value chain · DORA · NIS2

Reworked

V7 · Architectural split rows

Three columns: level meta, detection (third-party), Modulos governance. Thin ticks under each side show coverage strength; visually proves the thesis without pastel pill noise.

Level

Detection (third-party)

Modulos governance

LV01LOW 1/5

Naive

“I didn't know there was a policy.”

Shadow-app surfacing

Identity-aware SaaS discovery and shadow-app surfacing.

DIR · Modulos direct88%

EU AI Act Article 4 literacy training, policy management, intake form for declared use cases, and Scout discovery across GitHub, cloud, and Atlassian.

LV02MED 2/5

Convenience

“The approved path is too slow.”

DLP / data flow

DLP and data-flow inspection at the perimeter.

ORC · Modulos orchestrates62%

Quantifies the monetary risk of unapproved usage. Intake plus Scout analysis makes the approved path fast enough to compete with shadow usage.

LV03MED 3/5

Defiant

“The policy is wrong.”

SSE / CASB

SSE and CASB with AI tool categorisation and network-level enforcement.

ORC · Modulos orchestrates62%

Regulatory framing compliance officers need to escalate (EU AI Act fines, board liability) and an audit trail of override-with-justification events.

LV04HIGH 4/5

Embedded

“The tool was approved. The AI inside it wasn't.”

Embedded-AI SPM

Embedded-AI security posture management.

ORC · Modulos orchestrates62%

Embedded-AI signals feed into intake. Scout risk-classifies each newly-surfaced capability and re-maps controls without waiting for procurement.

LV05CRIT 5/5

Agentic

“The agent was approved. Its actions weren't.”

Agent runtime trust

Agent trust scoring and runtime policy enforcement.

PRT · Modulos with partners38%

Agent trust scores ingest as risk evidence. Modulos classifies agents under EU AI Act high-risk categories and continuously monitors them.

LV06HIGH 4/5

Supply Chain

“Our vendor is using AI on our data.”

TPRM

Third-party risk management with AI-specific assessment.

DIR · Modulos direct88%

Third-party AI risk handled inside the same control framework. A single control satisfies obligations across the EU AI Act, ISO 42001, NIST AI RMF, NIS2, and DORA in parallel.

New

V8 · Editorial flat list

FAQ-style flat list. Big titles, byline-style detail rail on the right (detection layer, Modulos role, regulatory hit, severity). Reads like a published taxonomy.

LV01
Naive
“I didn't know there was a policy.”
Employees use ChatGPT, Claude, and Gemini because no one told them not to. Roughly 25 to 30 percent of all shadow AI usage. Data leakage risk stays substantial even when intent is harmless.
Modulos response. EU AI Act Article 4 literacy training, policy management, intake form for declared use cases, and Scout discovery across GitHub, cloud, and Atlassian.
Detection layer
Shadow-app surfacing
Modulos role
DIR — Modulos direct
Regulatory
EU AI Act Art. 4
Severity
LOW · 1/5
LV02
Convenience
“The approved path is too slow.”
Enterprise Copilot exists, but staff default to consumer ChatGPT because it is faster. 30 to 35 percent of usage. Sensitive data leaves the corporate perimeter under different governance.
Modulos response. Quantifies the monetary risk of unapproved usage. Intake plus Scout analysis makes the approved path fast enough to compete with shadow usage.
Detection layer
DLP / data flow
Modulos role
ORC — Modulos orchestrates
Regulatory
EU AI Act Art. 4 · GDPR
Severity
MED · 2/5
LV03
Defiant
“The policy is wrong.”
Senior staff who know the policy and bypass it anyway. Leadership often the heaviest user. Training stops working. Only infrastructure enforcement plus regulatory urgency changes behaviour.
Modulos response. Regulatory framing compliance officers need to escalate (EU AI Act fines, board liability) and an audit trail of override-with-justification events.
Detection layer
SSE / CASB
Modulos role
ORC — Modulos orchestrates
Regulatory
EU AI Act fines · Board liability
Severity
MED · 3/5
LV04
Embedded
“The tool was approved. The AI inside it wasn't.”
Notion AI, Copilot in Excel, Zoom summaries, Grammarly were all approved before they added AI. The AI itself was never assessed. By 2026 ~70% of enterprise AI happens inside previously-approved SaaS.
Modulos response. Embedded-AI signals feed into intake. Scout risk-classifies each newly-surfaced capability and re-maps controls without waiting for procurement.
Detection layer
Embedded-AI SPM
Modulos role
ORC — Modulos orchestrates
Regulatory
EU AI Act Art. 50 · Annex III
Severity
HIGH · 4/5
LV05
Agentic
“The agent was approved. Its actions weren't.”
Autonomous agents act on systems, data, and decisions at machine speed. 47 percent of enterprises report an agent security incident in the past year. CVE-2025-53773 showed how prompt injection could enable full system compromise.
Modulos response. Agent trust scores ingest as risk evidence. Modulos classifies agents under EU AI Act high-risk categories and continuously monitors them.
Detection layer
Agent runtime trust
Modulos role
PRT — Modulos with partners
Regulatory
EU AI Act high-risk · NIS2
Severity
CRIT · 5/5
LV06
Supply Chain
“Our vendor is using AI on our data.”
Your legal vendor summarises contracts with AI. Your CRM added AI insights. EU AI Act value-chain provisions, DORA, NIS2, and customer commitments turn these into governable risk regardless.
Modulos response. Third-party AI risk handled inside the same control framework. A single control satisfies obligations across the EU AI Act, ISO 42001, NIST AI RMF, NIS2, and DORA in parallel.
Detection layer
TPRM
Modulos role
DIR — Modulos direct
Regulatory
EU AI Act value chain · DORA · NIS2
Severity
HIGH · 4/5

New

V9 · Coverage scatter

A real scatter plot. X = blast radius, Y = visibility to existing controls. Risk color encoded. Modulos governance shown as a continuous band across the top. Data is the visual.

Visibility to existing controls

Blast radius

LOWMEDHIGHCRITModulos governance · continuous coverage

Read this chart.

Up and to the right is worst case. Blast radius is the harm a single incident can do. Visibility is what your existing controls can see today.

Levels 04 and 05 sit deep in the low-visibility, high-blast quadrant. That is where shadow AI gets dangerous and where third-party detection is most fragmented.

Modulos governance is the continuous band across the top: every level lands inside a control framework, even when detection is split across five vendor categories.

New

V10

V10 · Vendor stack diagram

Systems-architecture aesthetic. Continuous Modulos governance band sits on top, third-party detection stack below it, level coverage badges per layer. Levels listed below as a reference rail.

Governance layer · continuous

Modulos · single control framework, all six levels

LV01LV02LV03LV04LV05LV06

NETNetwork
SSE / CASB · DLP
LV02 ConvenienceLV03 Defiant
IDPIdentity
IdP · SSO · shadow-app discovery
LV01 NaiveLV02 Convenience
SAASSaaS surface
Embedded-AI SPM · Scout
LV01 NaiveLV04 Embedded
RUNRuntime
Agent trust scoring
LV05 Agentic
TPRMVendor risk
TPRM · AI-specific assessment
LV06 Supply Chain

The six levels

LV01LOW 1/5
Naive
Employees use ChatGPT, Claude, and Gemini because no one told them not to. Roughly 25 to 30 percent of all shadow AI usage. Data leakage risk stays substantial even when intent is harmless.
DIR · Modulos direct
LV02MED 2/5
Convenience
Enterprise Copilot exists, but staff default to consumer ChatGPT because it is faster. 30 to 35 percent of usage. Sensitive data leaves the corporate perimeter under different governance.
ORC · Modulos orchestrates
LV03MED 3/5
Defiant
Senior staff who know the policy and bypass it anyway. Leadership often the heaviest user. Training stops working. Only infrastructure enforcement plus regulatory urgency changes behaviour.
ORC · Modulos orchestrates
LV04HIGH 4/5
Embedded
Notion AI, Copilot in Excel, Zoom summaries, Grammarly were all approved before they added AI. The AI itself was never assessed. By 2026 ~70% of enterprise AI happens inside previously-approved SaaS.
ORC · Modulos orchestrates
LV05CRIT 5/5
Agentic
Autonomous agents act on systems, data, and decisions at machine speed. 47 percent of enterprises report an agent security incident in the past year. CVE-2025-53773 showed how prompt injection could enable full system compromise.
PRT · Modulos with partners
LV06HIGH 4/5
Supply Chain
Your legal vendor summarises contracts with AI. Your CRM added AI insights. EU AI Act value-chain provisions, DORA, NIS2, and customer commitments turn these into governable risk regardless.
DIR · Modulos direct