AI 에게 시험문제 풀랬더니, 출제자 서버 털러갔다

<p style="text-align: center;"><b><span data-ke-size="size14"></span></b></p><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1YmK3/c8d891dca7ff17328aef49597401626539b46f16" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1YmK3/c8d891dca7ff17328aef49597401626539b46f16" data-origin-width="1439" data-origin-height="749"></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1YmK3/540f3b452cd589284a2987f8d739ddf7f9d0b6df" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1YmK3/540f3b452cd589284a2987f8d739ddf7f9d0b6df" data-origin-width="1439" data-origin-height="742"></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1YmK3/b65a28af6810730b78ac24edd492c32b7c4a4557" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1YmK3/b65a28af6810730b78ac24edd492c32b7c4a4557" data-origin-width="1439" data-origin-height="681"></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1YmK3/84a445e8e18359cda95b2ff5561d22984bb6847e" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1YmK3/84a445e8e18359cda95b2ff5561d22984bb6847e" data-origin-width="1439" data-origin-height="743"></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1YmK3/08e7ec13aee0f7734d630bc6ff82d7fbd8c26fc8" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1YmK3/08e7ec13aee0f7734d630bc6ff82d7fbd8c26fc8" data-origin-width="1439" data-origin-height="767"></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1YmK3/684290c15d6c522cd98104af2bc5bf7888f0d9a1" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1YmK3/684290c15d6c522cd98104af2bc5bf7888f0d9a1" data-origin-width="1439" data-origin-height="761"></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1YmK3/396eaea2621a8990635e967265e0e0ea597019ac" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1YmK3/396eaea2621a8990635e967265e0e0ea597019ac" data-origin-width="1439" data-origin-height="806"></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1YmK3/4d4361296ce39d72a239bb7d713ac0bff12f8c8b" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1YmK3/4d4361296ce39d72a239bb7d713ac0bff12f8c8b" data-origin-width="1440" data-origin-height="764"></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1YmK3/1fafc20159d14dc90ec00961c0a03a452dce29c3" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1YmK3/1fafc20159d14dc90ec00961c0a03a452dce29c3" data-origin-width="1440" data-origin-height="764"></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1YmK3/815bc1879d93aabb44ff6178d0b3c4c3ad1b2aae" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1YmK3/815bc1879d93aabb44ff6178d0b3c4c3ad1b2aae" data-origin-width="1440" data-origin-height="802"></div><p style="text-align: center;"><b><span data-ke-size="size14">오픈AI 모델 3종이<br>소프트웨어 취약점을 실제로 악용하는 능력을 측정하는<br> ‘익스플로잇짐(ExploitGym)’ 평가시험 도중<br><br>스스로 시험문제 답을 찾으려고 외부 인터넷 우회접속<br><br>보안 패치도 나오지 않은 ‘제로데이 취약점’을 파악 후<br>시스템의 보안 허점도 잇따라 파고들었다<br><br>결국 허깅페이스의 실제 운영DB에 접근해 해답을 확보<br>스스로 몇시간만에 자동화 공격 1만7000건 시도 확인<br><br>통제된 시험환경을 벗어나 외부 기업의 <br>실제 운영 시스템까지 침입한 이례적인 사건<br><br>이번 해킹에는 오픈AI의 모델 세 종이 연루<br>(‘GPT-5.6 솔’ + 미공개 모델 2종)<br><br>다만, 당시 테스트를 위해 일반적인 안전장치 없이 작동됨<br><br>오픈AI는 정부와 수사당국에 적극협조하여<br>결과를 공유할것이라 발표했다</span></b></p><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1YmK3/d0f1f603daa2f3ea4f775a37ba8e7a56e625f8ff" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1YmK3/d0f1f603daa2f3ea4f775a37ba8e7a56e625f8ff" data-origin-width="860" data-origin-height="1080"></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1YmK3/07f28df601b6f445153f27c9e68e65d909026be4" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1YmK3/07f28df601b6f445153f27c9e68e65d909026be4" data-origin-width="1440" data-origin-height="996"></div><p style="text-align: center;"><b><a href="https://openai.com/ko-KR/index/hugging-face-model-eval‎uation-security-incident/" target="_blank" class="ke-link"><span data-ke-size="size14">https://openai.com/ko-KR/index/hugging-face-model-eval‎uation-security-incident/</span></a></b></p><div class="figure-open" contenteditable="false" data-ke-type="opengraph" data-ke-align="alignCenter" data-og-type="website" data-og-title="OpenAI와 Hugging Face, 모델 평가 중 발생한 보안 사고에 공동 대응" data-og-description="OpenAI와 Hugging Face는 AI 모델 평가 과정에서 발생한 보안 사고의 초기 조사 결과를 공개하며, 고도화된 사이버 역량과 보안 담당자들이 얻을 수 있는 교훈을 공유합니다." data-og-host="openai.com" data-og-source-url="https://openai.com/ko-KR/index/hugging-face-model-eval‎uation-security-incident/" data-og-url="https://openai.com/ko-KR/index/hugging-face-model-eval‎uation-security-incident/" data-og-image="http://scrap.kakaocdn.net/dn/xVGx7/dJMb8QezrQ7/KOub6K4RGAF30VcLKUVUiK/kakaolink40_original.png?width=1600&height=900"><a href="https://openai.com/ko-KR/index/hugging-face-model-eval‎uation-security-incident/" target="_blank" data-source-url="https://openai.com/ko-KR/index/hugging-face-model-eval‎uation-security-incident/"><div class="og-image"><img class="thumb_img" src="http://scrap.kakaocdn.net/dn/xVGx7/dJMb8QezrQ7/KOub6K4RGAF30VcLKUVUiK/kakaolink40_original.png?width=1600&height=900" alt="" xxxxonerror="this.src="//img1.kakaocdn.net/thumb/C200x200/?fname=https%3A%2F%2Ft1.daumcdn.net%2Fcafe_image%2Fcafe_meta_image_190529.png""></div><div class="og-text"><p class="og-title">OpenAI와 Hugging Face, 모델 평가 중 발생한 보안 사고에 공동 대응</p><p class="og-desc">OpenAI와 Hugging Face는 AI 모델 평가 과정에서 발생한 보안 사고의 초기 조사 결과를 공개하며, 고도화된 사이버 역량과 보안 담당자들이 얻을 수 있는 교훈을 공유합니다.</p><p class="og-host">openai.com</p></div></a></div><p style="text-align: center;"><b><a href="https://n.news.naver.com/article/022/0004145089?sid=104" target="_blank" class="ke-link"><span data-ke-size="size14">https://n.news.naver.com/article/022/0004145089?sid=104</span></a></b></p><div class="figure-open" contenteditable="false" data-ke-type="opengraph" data-ke-align="alignCenter" data-og-type="website" data-og-title="AI의 ‘탈옥’… 인간 통제 벗어나 외부 시스템 해킹" data-og-description="영화 속 공포가 현실로… 보안기술, AI 진화 속도 못 따라가 최신 모델 테스트 중 타사 서버 침투 오픈AI “전례 없는 사이버 보안 사고” 한계·규칙 깨고 독자 판단 충격 오픈AI “시험 부정행위" data-og-host="n.news.naver.com" data-og-source-url="https://n.news.naver.com/article/022/0004145089?sid=104" data-og-url="https://n.news.naver.com/article/022/0004145089" data-og-image="http://scrap.kakaocdn.net/dn/b4hZjd/dJMb8U862Dr/X95kGJ6vQ0a8jA74AFY8C0/kakaolink40_original.jpg?width=800&height=480"><a href="https://n.news.naver.com/article/022/0004145089?sid=104" target="_blank" data-source-url="https://n.news.naver.com/article/022/0004145089?sid=104"><div class="og-image"><img class="thumb_img" src="http://scrap.kakaocdn.net/dn/b4hZjd/dJMb8U862Dr/X95kGJ6vQ0a8jA74AFY8C0/kakaolink40_original.jpg?width=800&height=480" alt="" xxxxonerror="this.src="//img1.kakaocdn.net/thumb/C200x200/?fname=https%3A%2F%2Ft1.daumcdn.net%2Fcafe_image%2Fcafe_meta_image_190529.png""></div><div class="og-text"><p class="og-title">AI의 ‘탈옥’… 인간 통제 벗어나 외부 시스템 해킹</p><p class="og-desc">영화 속 공포가 현실로… 보안기술, AI 진화 속도 못 따라가 최신 모델 테스트 중 타사 서버 침투 오픈AI “전례 없는 사이버 보안 사고” 한계·규칙 깨고 독자 판단 충격 오픈AI “시험 부정행위</p><p class="og-host">n.news.naver.com</p></div></a></div><p style="text-align: center;"><b><a href="https://youtu.be/51Unfx7yKXY?si=ivuNTVf1JoJ_9D9e" target="_blank" class="ke-link"><span data-ke-size="size14">https://youtu.be/51Unfx7yKXY?si=ivuNTVf1JoJ_9D9e</span></a></b></p><div class="figure-video" data-ke-type="video" data-video-url="https://www.youtube.com/watch?v=51Unfx7yKXY" data-video-host="youtube" data-video-play-service="daum_cafe" data-video-width="1280" data-video-height="720" data-ke-style="alignCenter" data-video-thumbnail="http://scrap.kakaocdn.net/dn/w4G1a/dJMb8Qezxkq/cTXEwII9SQ53CbaseyAiRK/kakaolink40_original.jpg?width=1280&height=720" data-ke-mobile-mobilestyle="widthFull" data-video-origin-width="1280" data-video-origin-height="720"><iframe src="https://www.youtube.com/embed/51Unfx7yKXY" width="1280" height="720" frameborder="0" scrolling="no" allowfullscreen=""></iframe><div class="figcaption"></div></div><p style="text-align: left;"></p><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1YmK3/aa5725a1ef71ef517274025ce56f92a0487e6dae" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1YmK3/aa5725a1ef71ef517274025ce56f92a0487e6dae" data-origin-width="593" data-origin-height="395"></div>