Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Toxic Oversight

Sponsored · Ship Features Fearlessly Turn features on and off without deploys. Used by thousands of Ruby developers.

Toxic Oversight

Quick privacy review among major LLM services, along with search for possible remedies. (OWASP Saitama MTG #25, talk #1)

Avatar for Takahiro Yoshimura

Takahiro Yoshimura

April 22, 2025
Tweet

More Decks by Takahiro Yoshimura

Other Decks in Technology

Transcript

  1. TOXIC OVERSIGHT OWASP SAITAMA MTG #25 TALK #1 Image by

    greenkozi on flickr, CC-BY-NC-ND 2.0
  2. TEXT BACKGROUND ▸ ChatGPT, DALL-E, Gemini, Copilot, Stable Diffusion, Midjourney..

    ▸ ͳΜͰ΋஌͍ͬͯΔ৺ڧ͍૬ஊ૬ख ▸ ͳΜͰ΋࡞ΕΔ΋ͷ͍͢͝ΞʔςΟετ ▸ ײ৘΋ײੑ΋͋Δ ▸ 2023/11/30ͷOpenAIʹΑΔChatGPTͷൃද͕ ൽ੾Γ… ΋͏2ࡀ Image by Xi on flickr, CC-BY-NC-ND 2.0
  3. TEXT BACKGROUND ▸ Large Language Models (LLM; େن໛ݴޠϞσϧ) ▸ ೖྗ͔ΒʮͦΕͬΆ͍ʯ͜ͱΛੜ੒

    ▸ σδλϧԽ͞ΕͨύϥϨϧϫʔϧυͷॅਓ ▸ 64-bit௨ΓͷseedͱͦΕ·Ͱͷݴ༿Ͱܾఆ࿦ తʹܾ·Δੈքઢ ▸ ୯ͳΔʮิ׬Ϟσϧʯ Image by Xi on flickr, CC-BY-NC-ND 2.0
  4. TEXT WHO I AM ▸ Takahiro Yoshimura (@alterakey) https://keybase.io/alterakey ▸

    Monolith Works Inc. Co-founder, CTO Security researcher ▸ ໌࣏େֶαΠόʔηΩϡϦςΟݚڀॴ ٬һݚڀһ (ʙ3݄)
  5. TEXT WHAT I DO ▸ Security research and development ▸

    iOS/Android Apps →Financial, Games, IoT related, etc. (>200) →trueseeing: Non-decompiling Android Application Vulnerability Scanner [2017] ▸ Windows/Mac/Web/HTML5 Apps →POS, RAD tools etc. ▸ Network/Web penetration testing →PCI-DSS etc. ▸ Search engine reconnaissance (aka. Google Hacking) ▸ Whitebox testing ▸ Forensic analysis
  6. TEXT WHAT I DO ▸ CTF ▸ Enemy10, Sutegoma2 ▸

    METI CTFCJ 2012 Qual.: Won ▸ METI CTFCJ 2012: 3rd ▸ DEF CON 21 CTF: 6th ▸ DEF CON 22 OpenCTF: 4th ▸ ൃදɾߨԋͳͲ DEF CON 25 Demo Labs (2017) DEF CON 27 AI Village (2019) CODE BLUE (2017, 2019) CYDEF (2020) etc. Image by Wiyre Media on flickr, CC-BY 2.0
  7. TEXT WOULD YOU LIKE TO HAVE A STRAWBERRY? ▸ 2024೥9݄ͷOpenAIͷGPT-4

    o1ͷެ։લ໷ ▸ strawberryͷrͷ਺Λ਺͍͑ͤͯͨ͞ΞΧ΢ϯ τ͕ݢฒΈౚ݁͞Εͨ… ͳͥʁ ▸ GPT-4 o1͸಺෦తʹ"strawberry"ͱݺ͹Ε͍ͯ ͨ… "r"Λےಓཱ͖ͯͯͪΜͱ਺͑ΒΕΔΑ͏ ʹ͢Δͧͱ͍͏ҙؾࠐΈ ※LLM͸͜ͷྨͷ៛ີͳਪ࿦͕େͷۤख ▸ ͭ·Γձ࿩಺༰Λݟ͍ͯΔ Image by TimAlosi.com on flickr, CC-BY-NC-ND 2.0
  8. TEXT ... ARE YOU OK? ▸ ʮChatGPTར༻ऀͷϝϯλϧ่յΛݒ೦ʯ →͍·Ͱ΋ܧଓతʹݟ͍ͯΔ ▸ ύεϙʔτِ଄͕Ͱ͖ͨͱ͍͏ใࠂޙɺ2࣌ؒఔ

    ౓Ͱີ͔ʹϓϩϯϓτ͕෧࠯͞Εͨࣄ݅ →΄΅ϦΞϧλΠϜʹݟ͍ͯΔ ▸ DALL-Eʹ͓͍ͯ΋ϓϩϯϓτ܏޲෼ੳ →ؙདͰ͢Αɺͱ Image by TimAlosi.com on flickr, CC-BY-NC-ND 2.0
  9. TEXT WE ARE WATCHING YOU, SIR ▸ ձ࿩಺༰͔Βࣗಈతʹ௨ใ͞ΕΔ͜ͱ΋͋Δ ▸ Meta

    AIͷγεςϜϓϩϯϓτʹʮෆ๏ߦҝ͕ ٙΘΕΔ৔߹ʹ͸ฦ౴ͤͣଈ࠲ʹ௨ใͤΑʯ ͱ͍͏هࡌ͕͋Δ ▸ ͲΕ͚ͩޮ͍͍ͯΔͷ͔͸෼͔Βͳ͍͕… llama͕πʔϧ΋࢖͑Δ͜ͱΛߟྀ͢Δͱͦ͜ ͦ͜ػೳ͍ͯ͠ΔՄೳੑ Image by telmo32 on flickr, CC-BY-ND 2.0
  10. TEXT YES, LOOKING AT YOU, SIR ▸ ؂ࢹ΍৘ใར༻͕Ճ଎͖ͯͨ͠ ▸ ྙཧنఆ΍ڝ૪༏Ґੑ͔Βͷཁ੥

    ▸ ChatGPTग़ݱॳ಄ͷ৘ใྲྀग़ٙ࿭ͱ͸ผ෺ →opt-outՄೳͩͬͨ →API͸ͦ΋ͦ΋ର৅Ͱ͸ͳ͔ͬͨ ▸ ͜Μͳ΋ͷ৴པͰ͖Δ͔ʁ →ʮѱ༻๷ࢭʯʮCSAMʯͱ͸ຐ๏ͷݴ༿ (·ͨOpenAI͸๷Ӵ࢈ۀೖΓ͍ͯ͠Δ) Image by loren chipman on flickr, CC-BY-NC 2.0
  11. TEXT TRUMP MADNESS ▸ ୈೋ࣍τϥϯϓ੓ݖ ▸ ࣗࠃ༏ઌओٛͱอ਎ ▸ શͯ͸ຽओౘͱҠຽͷ͍ͤ ▸

    ۙࢹ؟తํࡦ ▸ ൓DEI ▸ Ҡຽഉআ੓ࡦ →·͞ʹϑΝγζϜతͳ͜ͱ͕ى͖͍ͯΔ Image by IoSonoUnaFotoCamera on flickr, CC-BY-SA 2.0
  12. TEXT AT YOUR SERVICE, PRESIDENT ▸ ͍ΘΏΔΦϦΨϧώͷελϯυϓϨʔ͕ͻͲ͍ ▸ Өͷ੓෎ͱԽ͍ͯ͠Δײ৮ →ಛʹΠʔϩϯͱDOGE

    ▸ ൓DEI΁ͷศ৐͔ΒҰؾʹӈ܏Խ →ϩϏʔ׆ಈͷҰ؀ͱͯ͠ͷศ৐ʁ Image by dmoberhaus on flickr, CC-BY 2.0
  13. TEXT OUTFLOWING MADNESS ▸ ٸܹͳن੍؇࿨ํ޲ͷಈ͖ ▸ ੓࣏తѹྗΛ͔͚Δಓ۩ʹͳ͍ͬͯΔ ▸ ৘ใऩू͓Αͼར༻͸ଟ෼ଓ͘ ▸

    Ϟϥϧϋβʔυ͸݈ࡏ ▸ ࣃࢭΊ͸֎Ε͖͍ͯͯΔ (≒΍Γ͍ͨ์୊) Image by shioshvili on flickr, CC-BY-SA 2.0
  14. TEXT YES THEY ARE WATCHING US ▸ ϓϥΠόγʔ͕͔ͭͯͳ͍΄Ͳةݥʹ ▸ ͜Ε·Ͱ΋Ϋϥ΢υʹ͍ͭͯݴΘΕ͍͕ͯͨ

    →CSAMޡೝ໰୊ɺใ෮తใࠂ໰୊ ▸ ੜ੒AI͸ѱ༻๷ࢭͳͲͷҙຯ߹͍ͰαʔϏε؂ ࢹ͞Ε͍ͯΔ →Ͳ͜Ͱ΋ੜ੒AI…ʁ ▸ Google: ܉ࣄ΍؂ࢹʹར༻͠ͳ͍ͱ͍͏දه Λີ͔ʹམͱ͍ͯ͠ΔʢGemini…ʁʣ Image by ashley rose, on flickr, CC-BY-NC-ND 2.0
  15. TEXT YES THEY ARE WATCHING US ▸ ϓϥΠόγʔ͕͔ͭͯͳ͍΄Ͳةݥʹ ▸ xGrok

    3ʹมͳ͜ͱΛฉ͍͍ͯΔਓͨͪ… ▸ ࠃ಺ͷੜ੒AIαʔϏεʹมͳϓϩϯϓτΛ༩ ͍͑ͯΔਓͨͪ… ▸ ਎ͷةݥ͕ഭ͍ͬͯΔͱݴͬͯ΋ࢧোͳ͍ ▸ ࠓ͙͢ʹ΍Ίͨ΄͏͕ྑ͍ ʢ೔ຊͰ΋ཱͪೖΓௐࠪ͞ΕΔՄೳੑʣ Image by ashley rose, on flickr, CC-BY-NC-ND 2.0
  16. TEXT PRIVACY PLEASE, SIR ▸ ւ֎Ͱ͸໰୊ఏى͕ͳ͞Ε͖ͯͨ ▸ 9݄16೔ͷnatureͷهࣄ "Forget ChatGPT:

    why researchers now run small AIs on their laptops" ▸ ݚڀऀͨͪ͸Open-weightͷLLMΛࣗ਎ͷϚ γϯͰಈ͔͍ͯ͠Δͱ͍͏࿩ (Local LLM) ▸ techܥ͚ͩͰͳ͘ඇtechܥ΋ Image by calebjcook on flickr, CC-BY-NC-ND 2.0
  17. TEXT BECAUSE MY THOUGHTS ARE .. ▸ ೳྗ͸ྼΔͷʹͳͥʁ ▸ ࣗ਎ͷϚγϯͰ͋Ε͹׬શʹϓϥΠϕʔτ

    ▸ ୭ʹ΋௨ใ͞Εͳ͍ ▸ ৘ใΛར༻͞Εͳ͍ ▸ ࣗ༝ͳࢥࡧɾਪ࿦ɾ஌త࡞ۀ͕Ͱ͖Δ Image by Frankenstein on flickr, CC-BY-NC 2.0
  18. TEXT .. NONE OF YOUR BUSINESS ▸ ͞Βʹ͸ ▸ ֤ྖҬͷ஌ࣝਓ΍ݚڀऀ͕ɺӦརڝ૪ʹ૸Γ͸

    ͡Ί֤ͨࣾʹةػײΛ͍࣋ͬͯΔ͜ͱͷදΘΕ Image by Frankenstein on flickr, CC-BY-NC 2.0
  19. TEXT .. AM I WRONG? ▸ ೔ຊͰ͸Ͳ͏͔ ▸ NICT΍PFNΛ࢝Ίͱͯ͠Ϟσϧ͕࡞ΒΕ͍ͯΔ ͕…ͧͬͯ͜Closedɺग़དྷ͸……………

    ▸ ؂ࢹ͢ΔΤϯςΟςΟ͕੓෎΍௚׋ػؔʹ →͜Ε͸͜ΕͰ͍ͩͿେ͖ͳ໰୊ ▸ ࢖͏ͷͰ͋Ε͹Ϋϥ΢υ্ʹdeploy͢Δ͔ →Ͱ͖ͳ͚Ε͹ϋʔυ΢ΣΞ͝ͱങ͏͜ͱʹ →େมʹ࢖͍ʹ͍͘ Image by Toni Blay on flickr, CC-BY-NC-ND 2.0
  20. TEXT CLOSED MODELS ARE NOT YOUR FRIENDS ▸ Openͳͷ͸Swallow͘Β͍ ▸

    ೔ຊޠରԠͷ೉͠͞͸ΞυόϯςʔδͰ͸͋Δ ͷ͕ͩɺͦ͜ʹด͡ࠐ΋͍ͬͯΔΑ͏ʹݟ͑Δ ▸ ϘϥϯςΟΞҙࣝͷࠩͳͷ͔ɺͦΕͱ΋୯ͳΔ ৘ॹతن੍Λ༏ઌ͢ΔͨΊͩΖ͏͔ ▸ ೔ຊͷݚڀऀ͕೔ຊޠͰࣗ༝ʹߟ͑Δʹ͸͓ۚ Λࢧ෷Θͳ͚Ε͹ͳΒͳ͍…ͱ͸ɺ͏ʔΜ Image by WilsonB on flickr, CC-BY-SA 2.0
  21. TEXT ESCAPE PLAN A ▸ Local LLM؀ڥ: ࣗ਎ͷϚγϯͰinfer͢Δ ▸ ௕ॴ

    ▸ ׬શʹϩʔΧϧʹ༓ดͰ͖Δ ▸ ԿΛಈ͔ͯ͠΋໰୊ͳ͍ ▸ ୹ॴ: ▸ ܭࢉೳྗͱϝϞϦ͕ඞཁ ʢe.g. 8B:q8_0 -> 8.5GB, 70B:q4_K_M -> 43GBʣ ▸ Open-weightܥϞσϧͷΈ࢖༻Ͱ͖Δ
  22. TEXT ESCAPE PLAN A ▸ Local LLM؀ڥྫ ▸ ollama API

    ollama + Open-WebUI ▸ OpenAI-compatible API llama-cpp / vLLM + Open-WebUI llama-cpp / vLLM + LM Studio ... ※LM Studio͸proprietaryͰ͋Δ͜ͱʹ஫ҙ ▸ খن໛ͳΒllama.cppɺେن໛ͳΒvLLM͕༏ल (Flash Attention 2 vs. Paged Attention, ฒྻੑೳ)
  23. TEXT ESCAPE PLAN B ▸ ४Local LLM؀ڥ: ࣗ਎͕ॴ༗͢ΔΫϥ΢υܭࢉ؀ڥ ▸ ௕ॴ

    (※AWS) ▸ ASICͰͷinferͳͷͰߴ଎ ▸ ServerlessͰ࢖༻Ͱ͖ΔϞσϧ͕ଟ͍ (e.g. Claude, Deepseek R1, LLaMA 3.1 405B etc.) ▸ ୹ॴ ▸ ίετ͕ൃੜ͢Δ ▸ ϩʔΧϧʹ͸ด͡ΒΕͳ͍ Image by auxesis on flickr, CC-BY-SA 2.0
  24. TEXT ESCAPE PLAN B ▸ Ϋϥ΢υܭࢉ؀ڥͱ͸… ▸ ୯ͳΔߴ଎ͳܭࢉ؀ڥ →޷͖ͳϞσϧΛ࢖༻Ͱ͖Δ΋ͷ ▸

    APIαʔόͰ͸ͳ͍͜ͱʹ஫ҙ (cf. Azure vs. AWS) Image by auxesis on flickr, CC-BY-SA 2.0
  25. TEXT ESCAPE PLAN B ▸ ४Local LLM؀ڥྫ (※AWS) ▸ BedrockͰOpenAI-compatible

    APIΛग़͠… …ΫϥΠΞϯτ͔Βར༻ ▸ ͕ͩBedrock୯ମͰ͸OpenAI-compatible API͕ ग़ͤͳ͍ Image by auxesis on flickr, CC-BY-SA 2.0
  26. TEXT ESCAPE PLAN B ▸ bedrock-access-gatewayʂ ▸ AWS Sampleͱ্͕͍ͯͬͯ͠Δ΋ͷ https://github.com/aws-samples/bedrock-access-

    gateway/ ▸ BedrockΛ࢖༻͢ΔServerlessͳOpenAI- compatible APIαʔό࣮૷ɺͳͷ͕ͩ ▸ CloudFormationͰVPC͔ΒALB΍Βͳʹ΍Βશͯ ࡞ͬͯ͘ΕΔ… ͔͠΋1-clickϘλϯͰىಈͱ͔ ▸ େมʹ࢖͍ʹ͍͘୅෺
  27. TEXT ESCAPE PLAN B ▸ haversackʂ ▸ https://github.com/alterakey/haversack ▸ SAMͰॻ͖௚͠ɺ͔ͭ༨ܭͳϦιʔε࡞੒Λ

    ഉͨ͠΋ͷ ▸ ECS༻ͷΠϝʔδΛϕʔεʹ͢Δ͜ͱͰ ▸ ALB΋API Gateway΋ෆཁʹʂ ʢ୯ҰͷLambda Function URLͰఏڙʣ Image by avlxyz on flickr, CC-BY-SA 2.0
  28. TEXT ESCAPE PLAN B ▸ haversackʂ ▸ bedrock-access-gatewayͷECR΁ϩάΠϯޙ sam build,

    sam deploy -gͰߏஙdone ▸ ͪͳΈʹhaversackɺͱ͸ AD&Dʹొ৔͢Δຐྗ͕͔͔ͬͨbagͷҰछ Image by avlxyz on flickr, CC-BY-SA 2.0
  29. TEXT ESCAPE PLAN B ▸ ४Local LLM؀ڥྫ (※AWS) ▸ Bedrock

    + haversack + Open-WebUI ▸ Bedrock + haversack + LM Studio ▸ Bedrock + haversack + etc.. ▸ શͯserverlessʂ
  30. TEXT CAVEATS ▸ Ϟσϧͷઃఆ஋ʹ͍ͭͯ (e.g. System Prompt, context size, temp.

    etc.) ▸ ެࣜAPIʹ͓͚Δ૝ఆͱɺOpen-WebUI΍LM StudioͳͲͷҰൠతΫϥΠΞϯτ͕૝ఆͯ͠ ͍Δ஋ͱ͸͔ͳΓҧ͏ ▸ ࠷େݶʹ׆༻͢Δʹ͸୳ࡧ͢ΔΑΖ͠ Image by Aleksey Gureiev on flickr, CC-BY-NC-ND 2.0
  31. TEXT TAKEAWAYS ▸ ೔ຊʹ͓͚ΔLocal LLMͷར༻ʹ͸ →Open-weightͳϞσϧ͕଍Γͳ͍… →ਅͷҙຯͰࣗ༝ͳࢥࡧ͸Ͱ͖ͳ͍ ▸ ४LocalͰ͋Ε͹·ͩͳΜͱ͔ →ͨͩίετ͸͔͔Δ

    ▸ OpenAI-compatible API͔ollama APIΛग़ͤ͹ ༷ʑͳΫϥΠΞϯτ͔Βར༻ՄೳʹͳΔ Image by Chris Randall on flickr, CC-BY-NC 2.0
  32. TEXT ONE MORE THING ▸ Deepseekࣄ݅ ▸ ৘ใྲྀग़͕ͱ͔ར༻ن੍Λͱ͔ͳΜͱ͔૽͕Ε ͍ͯΔ͕ɺ͋͘·Ͱ΋ΞϓϦ΍ެࣜαΠτͷ࿩ ▸

    ୯ʹinfer͢Δ͚ͩͰ͋Ε͹ԿΒӨڹͳ͍ →४Local؀ڥͰ͸҆͝৺Λ ▸ Open-WebUI͸Webϕʔε… ͱ͍͏͜ͱ͸ϞόΠϧ୺຤͔ΒͰ΋ར༻Մೳʂ →ΞϓϦྨ͸Ұ੾ෆཁ Image by dakohuang on flickr, CC-BY 2.0