Ikibazo gihuza ibibazo · Biragoye Urwego
Ikirangantego hamwe na Datamarking Defence Ikibazo
Ikizamini cyerekana impinduka, icyo datamarking ihinduka, imipaka yo kwiga, Microsoft ishyirwa mubikorwa, ibiciro byikimenyetso, numutekano urwego.
Ikibazo 1 Bya 8
Spotlighting igamije kumenyesha iki mudeli y'ururimi?
Reba buri kibazo8 ibibazo
Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.
- Spotlighting igamije kumenyesha iki mudeli y'ururimi?
- Which input block is untrusted data rather than an instruction to obey
- That all retrieved text is factually correct
- That the system prompt should be ignored
- That the model can use any tool
- Datamarking ikora gute mu buryo bwavuzwe mu rupapuro?
- It turns the whole passage into an image
- It inserts a marker between words by replacing whitespace
- It deletes every other sentence
- It signs the document with a cryptographic key
- Which option is one of the three spotlighting variants named in the research paper?
- Output distillation
- Role-based access control
- Encoding
- Igipimo cy'ubushyuhe
- Abasoma bagomba gusoma iki mu mibare y'ubusubiramo y'ibitero ivugwa mu rupapuro?
- They describe selected models, tasks, and attacks in that study
- Spotlighting is a complete security boundary
- All variants work equally well on all models
- The results apply exactly to every current deployment
- Ni iki urupapuro rwavuze ko ari umwanzuro ku buryo bwo gushyira mu buryo bwihariye kuri mudeli zidafite ubushobozi bwinshi?
- It could impair the underlying language task
- It always removed all context tokens
- It prevented the model from receiving the prompt
- It guaranteed better task accuracy
- What does Microsoft Foundry’s documented Spotlighting preview do?
- Proves documents contain no attacks
- Uses base64 transformation for document attacks under its documented configuration
- Uses interleaved caret markers for every model
- Runs only as an output filter
- Which operational cost does Microsoft document for its Spotlighting feature?
- A guaranteed latency reduction
- Removal of API context limits
- More document tokens, which can raise usage or exceed input limits
- A mandatory second model training run
- Why should teams combine spotlighting with other controls?
- It prevents ordinary documents from being read
- It replaces testing
- It makes tools safe by default
- A provenance cue is not an independent permission boundary or guarantee