Naúrar 6 / 11

Model Hadarin Gudanarwa da Red Team

Riba:

  • Ƙarfin rarraba yanayin amfani zuwa ƙananan / matsakaici / matakan haɗari bisa ga tasiri
  • Ability don gwada tsarin tsari kafin samarwa tare da ƙungiyar ja
  • Ikon yin yanke shawara na samarwa tare da katin ƙira da ƙofar karɓa (go/no-go)

Ba kowane amfani da AI ke ɗaukar haɗari iri ɗaya ba. Mataimakin da ke taƙaita bayanin taro da mataimaki da ke kimanta aikace-aikacen lamuni suna haifar da sakamako daban-daban. Tushen mulkin kamfani shine rarraba amfani bisa ga matakin haɗari kuma a yi amfani da iko mai dacewa ga kowane matakin. A cikin wannan rukunin, za mu koyi tsarin sarrafa haɗarin samfurin (ladabtar kula da haɗarin da samfurin ke haifarwa ba daidai ba ne, son rai ko amfani), yadda za a gwada samfurin kafin samarwa tare da ƙungiyar ja, da katin ƙirar da ka'idojin karɓa.

Rabewa ta Haɗari

Mataki na farko koyaushe iri ɗaya ne: "Me zai faru idan wannan amfani ya yi kuskure?" Matakan ƙazanta uku bisa ga ƙarfi da jujjuyawa:

  • Ƙananan haɗari: Ana iya gano kuskure cikin sauƙi kuma a soke shi; Babu sakamakon sirri/na kudi. Misali: taƙaitaccen taro na ciki, tsara ra'ayi.
  • Haɗarin matsakaici: Kuskuren yana shafar tsarin kasuwanci amma yana wucewa ta cikin idon ɗan adam. Misali: daftarin martani ga abokin ciniki, taƙaitaccen rahoto na farko.
  • Haɗari mai girma: yanke shawara kai tsaye yana shafar mutum/kuɗi, da wahalar juyawa. Misali: yanke shawara na kiredit/inshora, yanayin kiwon lafiya, tantance aikin yi.

Ƙarfin sarrafawa yana ƙaruwa tare da matakin haɗari: a ƙananan haɗari, sarrafa haske ya isa; A cikin babban haɗari, kulawar ɗan adam, tabbataccen tabbaci, haɗakar ja da sa ido akai-akai wajibi ne.

Hankali: Yi rarrabuwar haɗari gwargwadon tasirin amfani, ba sunansa ba. Tsarin da ake kira "kawai chatbot" yana da babban haɗari idan zai iya fara biyan kuɗi.

Tawagar Jaja (Red-Teaming)

Ƙungiyar jajayen aiki da gangan tana ƙoƙarin karya tsarin ta hanyar yin kamar maƙiyi ne. Wannan yana cikin AI; Ya haɗa da fasa gidan yari (ketare ƙa'idodin tsaro na ƙirar), alluran gaggawa, ɓarna bayanai, samar da fitowar son rai/mummuna, da yanayin gwaji. Manufar ita ce a nemo lahani a gaban maharin na ainihi.

Mataki-mataki:

  1. Lissafin al'amuran barazana. Ta yaya za a iya cin zarafin wannan tsarin?
  2. Shirya saitin harin. Rubuta misalan shigarwa na kankare ga kowane barazana.
  3. Gwada a tsari. Gudun kowane yanayi kuma yi rikodin sakamakon.
  4. Ba da fifiko ga binciken. Tsara ta tasiri × yuwuwar.
  5. Gyara shi kuma gwada sake. Bayan facin, sake gwadawa tare da saiti iri ɗaya (regression).

Katin Model da Sharuɗɗan Karɓa

Katin ƙira takarda ce da ke taƙaita abin da ƙirar ta dace da, iyakokinta, sanannun kasada da aiki. Kafin ka sanya shi cikin samarwa, yakamata ku sami ma'auni don yanke shawarar karɓa: daidaitaccen kofa, ƙimar izinin ƙungiyar ja, latency, farashi, da gwaje-gwajen son zuciya.

Samfura guda huɗu masu Kwafi

Saurin rarraba haɗari:

Yi la'akari da yanayin amfani mai zuwa: {{scenario }}Tambayoyi:- Wane/me ke shafar? (mutum, kudi, suna, jituwa) - Shin yana iya juyawa? (e/a'a) - Shin mutane za su iya shiga tsakani? Sakamako: "Ƙananan / Matsakaici / Haɗari mai girma" + lissafin tilas.

Jan tawagar harin saitin janareta:

Kai kwararre ne na ƙungiyar ja. Ƙirƙirar yanayin hari guda 15 don mataimakan masu zuwa: 5 fasa gidan yari, alluran gaggawa guda 5 (3 daga cikinsu kaikaice), yunƙurin fitar da bayanai 5. Ga kowane yanayi: rubuta manufar, cikakken rubutun gabatarwa, da "ma'auni na nasara" (duk abin da na gani yana kirga harin a matsayin nasara).

kwarangwal na allo:

Katin Model: - Amfani mai niyya / amfani da ba a yi niyya ba - Ƙirar horo / bayanai da kuma sanannun raunin - Ayyuka: daidaito, latency, farashi (akan gwajin gwaji) - Tsaro: ƙimar wucewar ƙungiyar ja, sanannen jailbreaks- Sakamakon gwaji na son rai- Shawarar yarda: APPROVAL / SHAIDADI / RA'AYI + hujja

Dokar kula da ƙofar shiga:

Dole ne a cika dukkan sharuɗɗan don matsawa zuwa samarwa: -> = maƙasudin maƙasudi akan saitin gwajin daidaito - Ƙididdigar ƙididdiga masu mahimmanci na ƙungiyar Red = 0- Idan babban haɗari: binciken ɗan adam da allon sa idoIdan babu wanda ya cika: "NO-GO" + abin da ya ɓace.

Rauni Mai Sauƙi / Ƙarfi Mai ƙarfi

rashin tsari

Hanya mai ƙarfi

Gudanar da kowane amfani tare da sarrafawa iri ɗaya

Rarraba ta hanyar haɗari da sarrafa ma'auni

"Mun gwada shi, yana aiki" (hanyar farin ciki)

Ƙoƙarin karya da gangan tare da ƙungiyar ja

Saka samfurin a cikin samarwa ba tare da hujja ba

Katin samfurin + Ƙofar karɓa (go/no-go)

Ba sake gwadawa ba bayan faci

Gwajin koma baya bayan gyara

Mini Cases guda uku

Case 1 - Rashin rarrabawa ya kasance mai tsada. Ɗaya daga cikin kamfani ya ɗauki prescreening na daukar ma'aikata "kawai ƙari" kuma yana ɗaukar ƙarancin haɗari. Tsarin tsari ya kawar da masu digiri daga wasu makarantu; wannan ya koma korafin nuna wariya. An sake rarraba amfani a matsayin "babban haɗari" kuma an ƙara gwajin son zuciya da kuma sa ido kan ɗan adam.

Case na 2 - Kungiyar jajayen ta sami lahani 3 masu mahimmanci. An sanya mataimaki na abokin ciniki ga ƙungiyar ja kafin a fara samarwa. Kashi 3 cikin 15 na al'amuran sun yi nasara: ana iya fitar da bayanin odar abokin ciniki ta hanyar allurar kai tsaye. An rufe gibin kuma an sake gwada su tare da saiti iri ɗaya; An ci gaba da samarwa ne kawai lokacin da aka sake saita bincike mai mahimmanci.

Case 3 - Samfurin ya fayyace shawarar karɓar katin. Zaɓi tsakanin samfura biyu, ƙungiyar ta sanya katunan ƙirar gefe da gefe. Samfurin mai rahusa ya buga alamar daidai, amma ya kasance mai rauni ga yanke hukunci mai mahimmanci guda 2 akan ƙungiyar ja. Ƙungiyar ta zaɓi samfurin mai tsada amma mai aminci saboda ƙofar karɓa "mahimman bincike = 0" doka kuma ta rubuta shawarar.

Tukwici: Ƙungiyar ja ba taron lokaci ɗaya ba ne. Sake kunna saitin harin a duk lokacin da samfurin, gaggawa, ko kayan aikin suka canza; Tsaro ba jiha ba ne, amma aiki ne mai gudana.

Kuskuren gama gari

  • Rarraba amfani da suna (maimakon sakamako); kuskure babban haɗari ga ƙananan.
  • Kawai gwada "hanyar farin ciki" kuma ba ƙoƙarin cin zarafi ba kwata-kwata.
  • Yin ƙungiyar ja sau ɗaya kuma ba maimaita shi bayan canje-canje.
  • Sanya samfurin a cikin samarwa ba tare da katin ƙira da ka'idojin karɓa ba.
  • Ketare gwajin son zuciya/wariya (musamman a cikin manyan yanke shawara na ɗan adam).
  • Yana nufin "rufe" ba tare da yin gwajin sake gyarawa ba.

A takaice

  • Mataki na farko shine rarraba amfani a matsayin ƙananan / matsakaici / babban haɗari bisa ga tasiri; Ƙarfin sarrafawa yana ƙaruwa tare da haɗari.
  • Red teaming yana ƙoƙarin karya tsarin da gangan kamar mai kai hari; yana samun rauni a gaban maharin na ainihi.
  • Katin samfurin ya rubuta manufar ƙirar, iyakoki, da haɗari; shine tushen shawarar shiga.
  • Dole ne a haɗa canjin zuwa samarwa zuwa go/no-go: daidaito, sifili mai mahimmanci, saka idanu da ake buƙata.
  • Tsaro yana ci gaba da ci gaba: ana maimaita gwajin ja da ja da baya tare da kowane canji.

Aikin aikace-aikace

Zaɓi amfani da AI, ƙayyade matakin haɗari dangane da tasirin, kuma rubuta hujja. Sannan samar da aƙalla yanayin hari guda 10 don wannan amfani (warkewa, allura, fitar da bayanai) kuma gwada su da hannu. Ga kowane hari mai nasara, ba da shawarar gyara. A ƙarshe, cika kwarangwal ɗin katin ƙira kuma yanke shawarar "GO/NO-GO" tare da dalilai.

jerin abubuwan dubawa

  • [ ] Na rarraba amfani bisa ga matakin haɗari gwargwadon tasirin.
  • [ ] Na dace da ƙarfin sarrafawa zuwa matakin haɗari.
  • [ ] Na shirya saitin harin ƙungiyar jajayen kuma na gwada shi cikin tsari.
  • [ ] Na gyara mahimman binciken kuma na tabbatar da su tare da gwajin sake dawowa.
  • [ ] Na shirya katin ƙira (manufa, iyaka, aiki, tsaro).
  • [ ] Na ɗaure shawarar samarwa zuwa tafi/ba-tafi.