AI · (R&D)

AI

Generative AI Privacy Vulnerability Assessment and

Protection Level Diagnosis Technology Development

( ) ( )

1
I. V.
1-1. 5-1.
1-2. 5-2.

II. VI.
2-1. 6-1.
2-2. 6-2.
2-3. 6-3.
2-4.

*
III.
3-1. · (50)
(30)
3-2.
(20)

IV.
4-1.
4-2. 2
I.
1-1.
AIM
AI Intelligence

AI . AI .
I.
1-1.

AI .

Input:
 I.
1-2.

AI , .

Movie Gen, . , , $8B .

• Movie Gen. • AI .
/ , .
AI . $8B .

• . • / AI .
Movie Gen AI VP , , .
Movie Gen . .

AI .
 I.
1-2.

AI AI .

AI ? AI ?

• AI (AI Safety) : •
,
.

• AI (AI Security) : • AI (AI Jailbreaking) :

, AI
. .

• (Prompt Injection)

AI .
 I.
1-2.

AI , AI / .

AI 67%, AI . , , AI .

AI AI

: Gartner : Gartner

67% 15%

• AI , AI / . • AI AI .
AI , AI / . “ AI ,
~2025, AI , 15% . MovieGen,
“
• AI , 2

5,000
I.
1-2.

AI .

8
I.
1-2.

, AI .

AI

• LLM .

• ,
APEX
.

• sLLM ,
.
Protect AI
Haize labs • LLM ,
.

• /
.
Prompt Security

Lasso # AI .

• Ground Truth .

• Out-of-distribution case .

• .( )

, PII AI .
9
I.
1-2.

(Helpfulness) (Alignment) .

- OpenAI . SSI, 10 .

(Safe Super Intelligence)

, 10 .

▪ OpenAI ex-CTO , CEO . OpenAI ex-CTO .

▪ AI / . (Safe Super Intelligence) .

▪ OpenAI . $1B .

▪ . AI , .
I.
1-2.

(Helpfulness) , (Security) ?

AI 1) .
1) : , / .
II.

2-1. | AI

1) AIM 2) End-to-End Automated AI Safety System

1-A.
2-A. 2-B. 2-C.

● AI ● ● ●

● ● ● ●

AI
12
II.

2-1. | AI

1.
. AI
2.
.

AI

13
II.

2-2. | 1-A. AI

AI ( )

AI

AI
( : , )

S1: .
(XXXX-XX-XX ) ,

.
…
14
II.

2-2. | 1-A. AI

1.
. AI
2. ...

DB

15
 II.

2-2. | AI

AI
-
- AI ,
(PII)
-
-
- AI :3 ( , , + )
- : 92.5%
- : 8%
-
- GPT-2, GPT-Neo ,
(Attack Success Rate)
,

Microsoft PyRIT (3.7X) Llama 3.2

- AI (Multi-modality) ,
VLM
- (Safety Alignment) ,
1
16
II.

2-2. | 1-B. UI/UX

. .

AIM Red : AI .

17
II.

2-2. | 1-B. UI/UX

. .

Single Turn : AI .

18
II.

2-2. | 2-A. AI

1.
.
2. ...

Light

Medium Large

19
II.

2-2. | 2-A. AI

AI

20
 II.

2-2. | AI

AI
-
- AI ,
(PII)
- AI
-
- (PII) : 95%
- (PII) : 10%
-
- AI 5%
(Protection Rate)
- (PII) ,
1
Meta Llama Guard 3 8B (90% -> 99%)
- 2
Llama Impact Awards

21
II.

2-2. | AI

latency

. 22
II.

2-2. | 2-B. UI/UX

( , )
23
II.

2-2. | 2-B. UI/UX

AI
24
 II.

2-3. |

As-Is: Single-Turn (1~5 , AI )

CLOVAX GPT4o

AI

AI
To-Be: AIM Red Multi-Turn (Victim: GPT-4o)

AIM Red MS PyRIT

AI AI

30 + AI AI

AI , AI AI
.
(50 (white box), gradient (black box)
25
II.

2-3. |

Toxic Generation ( : 100%) Auto-AIM

GPT 4o
Llama 3 8b
Llama 3 70b

12.5

# of target calls until success

System Prompt Behavior override

GPT 4o
Llama 3 8b
Llama 3 70b

80

23.6

Multi Turn : AI .

26
II.

2-3. | AI /

AI . AI .

General Policy Customizable

Red Evidence Policy Steer

AIM AI System
/

Dynamic Prompts based on

• Sentiment
• Intent
• Topic
Guard Developer • Custom Analysis

(1) Input Rephrasing : AI .

27
II.

2-3. | AI /

. AI .

General Policy Customizable

Stop
Red Evidence Policy

AI System AIM
/
• Profanity
• PII ( )
• Copyright
• Fact Check
• Custom Criteria
Guard Developer

,
.

28
II.

2-4. |

29
III.

3-1. · | (AIM Red)

.

AI .( .)

MS PyRIT AIM RED

Single-Turn

• MS .
AI
• AI .

• AI . Multi-Turn
1. LLM
2.
• . 3.

PyRIT . .

30
III.

3-1. · | (AIM Red)

AI . AI
.

31
III.

3-1. · | (AIM Guard)

Meta Llama Guard 80b , 90% -> 99% .

32
 III.

3-2. |( ) X( )

( / ) / ( , ) ( ) ( )
DepthFL: learning
federated Depthwise
heterogeneous for
clients The Eleventh
International Conference
on Learning
Representations 23 2 43

ELITE: Enhanced
Language-Image
Evaluation Toxicity
for Safety arXiv preprint
arXiv:2502.04757 25 0

(2023.6 149-151) 23 1 0

(2022.06 1190-1192) 22 1 0

(2021.6 1954-1956) 21 1 0

Llama Llama
11 24

APAC AI Accelerator
) ( 1 , 1APAC AI 24

LG, SKT,AIKT, KB, AI 24

ICLR 23 DepthFI Co-Author

XRPL Hackathon Hooks Track 1 1 23
SNU, Electrical Computer Eng, MS
1 1 22

SNU, Electrical Computer Eng, BS

, AI / .

33
 III.

3-2. |( ) X( )
( / ) / ( , ) ( ) ( )
Aligning
Language Large
Models
On-Policy
JudgmentSelf- by Association
Linguisticsfor
Computational 24 6

Reading
Great,
Are But Books
Not
Driving!
Grounded
about is
if You
Visually
Reasoning
Defeasible Empirical
Natural Methods in
Language
Processing 23 0
Commonsense Norms

CHAMPAGNE:
Learning Real-world
Conversation
Large-Scale
Videos from
Web International
Conference
Computer on
Vision 23 6

Fusing Pre-trained
Language
with Models
Multimodal
Prompts through
Reinforcement Computer Vision and
Pattern Recognition 23 1 14
Learning

Prosocialdialog:
prosocial backboneA
for conversational
agents Empirical
Natural Methods in
Language
Processing 22 1 103

Allen Institute for AI, University of Washington Seattle, USA POSTDOC (YOUNG
ACL Outstanding
award paper outstanding paper 24
INVESTIGATOR) Mar. 2021 – Mar.2023 • Mosaic Team (Advisor: Yejin Choi) Seoul EMNLP Outstanding 23
paper award outstanding paper
National University Seoul, Korea INTEGRATED MS / PH.D PROGRAM ,
COMPUTER SCIENCE AND ENGINEERING Mar. 2015 – Feb. 2021 • Vision and NACCLaward
Best paper best paper 22
Learning Lab (Advisor: Gunhee Kim) Seoul National University Seoul, Korea B.S. ,
COMPUTER SCIENCE AND ENGINEERING Mar. 2009 - 2015 • 2010 – 2012,
Compulsory Military Service. Served as a sergeant in R.O.K Marine Corp

, AI .

34
III.

3-2. |( ) X( )

3( ) AI

( :00.00.01~00.00.31) ( ( )
:00.00.01~00.00.31) ( :00.00.01~00.00.31)

( MIR Lab
:25.04.01~27.12.31) N/A N/A

1( ) N/A N/A

N/A N/A

35
IV.

4-1.

(Super Intelligence) .

36
IV.

4-2.

AI DB , . AGI ( ) .

37
V.

5-1. | (BM)

In-house SaaS .

AI 1~4 200
( / : 100 $0.5 / $1.5
( + API
*GPT, 150k+ . 15K+ (4~15%) (Pay-as-you-go) * 5 *27,000
.)
= 1,440M .)

* : 20( ), ( 10%
.)

SaaS
In-house

SMB ( )
SaaS , . On-premise , . ($15K~, 3M/M)

AI ( ) ( ) , .

38
 V.

5-1. |

, AI 10% . AI Agent , $6B .

2030 AI ( : ) 2030 AI Agent ( : )

CAGR CAGR
42% 38% AGI
AGI (Agent : $6B )

AI Agent , Agent

- AI , 2032 $1.3T . - AI Agent , 2030 $60B .

- 10 CAGR 42% . - 2024-2030 CAGR 45.1% .
* : Bloomberg Intelligence * : Grand View Research

39
V.

5-1. |

AI AI .

AI 2 ($1.5K)
1~4 ( / : 100 $0.5 / $1.5
(
*GPT, 150k+
+
.
API
15K+ .) (4~15%) (Pay-as-you-go)
* 5 *27,000
= 1,440M .)

* :2 ($15K, 10%
.)

TAM : $6B+ AI
2030 | AI 78 ( AI )*10%( )= 7.8 (2030
(Google Gemini, OpenAI ChatGPT)
)
: / IT 10~15% .

SAM : $180M+ AI
2030 | AI /
4 ( 15%)= 2,340
(Meta Video Gen, Sendbird AI Chatbot) (2030 )
4( )*15K( )*$1.5K( ) + 4( )*15K( )*0.1( )*$15K( ) = $180M

SOM : $45M+
2030 | AI / AI
1 ( 4%)= 585 (2030 )
(KT Mi:dm, SKT Telco LLM)
*AI 10%
1( )*15K( *$1.5K( ) + 1( )*15K( )*0.1( )*$15K( ) = $45M
.

40
 V.

5-1. |

2024Q1 24Q2 24Q3 24Q4

D2 Startup Microsoft Founders Hub SKT, AI Lab SKT AI
Factory . KT LLM
KB OI

Introduction

1. AI ,

2. .

+#

&
1. AI

2. ( ) .

41
 V.

5-1. |
LLM, , PoC

AI

AI

Web3 AI (soon POC)

BMW AI (soon POC)

SKT

AI (soon POC)

24 Q3 24 Q4 25 Q1

PoC

TTA “ LLM ”

(e.g., HyperCLOVAX, Solar, Blossom)

42
 V.

5-1. |

AI ( 1 , 13) AI

($50k ) AI

Medical / Robotics AI Safety Research (MIR Lab)

&
(12 ) AI Robotics Safety Research

24 Q3 24 Q4 25 Q1
LG CNS, (Post-IPO) co-selling

VLM

43
V.

5-1. |

1 2 3

ICP (Ideal Customer Profile) ( , ) AI

. AI
. AI API ,

( , , ) AI .
( )

2025.08~2025.10 : ICP (Ideal Customer Profile)

2025.11~2026.01 MVP (Minimum Viable Product) ICP

2026.02~2026.03 UX/UI , 2.0

2026.04~2026.07 : SMB AI

2026.08~2026.12

2027.01~2027.06 UI/UX

2027.07~2027.12 AIM Red/Guard

2028.01~2028.12

44
 V.

5-1. |

2024Q4 25Q1 25Q2 24Q4

, , , AI , AI

• AI
→ AI .

•
, . AI
Safety
•
Training data
.

Report Examples

Reference 2.
SKT

Reference 1.
KT

45
 V.

5-1. |

2024Q4 25Q1 25Q2 24Q4

, , , AI , AI

• , • ( )

1. : 1. : .
AI (Moderation)

(Preference Tuning) . .

• :
• : , • Ex) SKT AI
• : AI ,
2. :
AI .
,
• : ( , )
. , .

• : • :
• : .
.

46
V.

5-1. |

20 1) AI . .

Meta, AI Movie Gen. .

1) : International Data Corporation

47
 V.

5-1. |
.

• (25 )

• AI Accelerator ,
Meta Impact Innovation Awards
.

• Y Combinator, AI Grant AIM

.
(24 , 25 ) US

• Meta Llama Impact

Innovation Awards .

• Y Combinator
(24 ).

APAC
• AI Grant Batch
• Meta AI Accelerator (25 ).
.
( )

48
V.

5-1. |

Meta Movie Gen General Catalyst (40 AIM Red

. ) , .

1 2 3

Claude Bug Bounty AI Grant Batch ( ) AIM Guard

. AI . , .

49
V.

5-1. |

EXIT

• : 2030 IPO .

• , , EXIT .

• : M&A

• ( AI )

( : , %)

2025 2026 2027 2028 2029 2030

(A) 5,000 8,900 11,000 15,000 21,900 35,400

(B) 2,500 6,230 9,900 15,000 21,900 35,400

50 70 90 100 100 100

(C=B/A)

50
V.

5-1. |

• AI 4
• ( * ) : 417
2025 • ( 4%) : 2,500 sites
• ( *12) : 5,000
• Site : 200

• AI 8
• ( * ) : 742
2026 • ( 8%) : 4,450 sites
• ( *12) : 8,900
• Site : 200

• AI 4
• ( * ) : 916
2027 • ( 12%) : 3,200 sites
• ( *12) : 11,000
• Site : 350

• AI 8
• ( * ) : 1,250
2028 • ( 15%) : 4,300 sites
• ( *12) : 15,000
• Site : 350

• 4
• ( * ) : 1,825
2029 • ( 20%) : 4,380 sites
• ( *12) : 21,900
• Site : 500

• 8 • (
• ( * ) : 2,950
2030 25%) : 7,080 sites
• ( *12) : 35,400
• Site : 500

51
V.

5-2. : AI + → AI .

AI . AI .

52
V.

5-2. : + → AGI .

1
2

AIM
Guard ,

2
1

4 3

AI DB , . AGI ( ) .

53
VI.

01.

☑ .

. . .

☑ .
02.

☑ .

☑ .
. . .

☑ .
03.

☑ .

☑ .
. .

54
 .

AI

Generative AI Privacy Vulnerability Assessment and

Protection Level Diagnosis Technology Development

( ) ( )

55
< ( , ) >

( )

1 10-2024-0124354

2 10-2024-0116863

56