Whitelist
A filtering mechanism that permits only pre-approved sources to pass, operating on a "default deny" principle. It offers high security but low flexibility.
Timeline (last 90 days)
Because models cannot fundamentally distinguish "data" from "instructions", prompt-based protection alone is unreliable; a viable approach is to have the model only output suggestions, with all actions that have side effects executed by application code according to a whitelist, and to perform schema validation and permission checks on model output.
A single-sender inbox essentially turns the entire inbox into a whitelist: by default it rejects all unknown sources, and only explicitly added senders can deliver.
A whitelist is a filtering mechanism that only allows pre-approved sources to pass, as opposed to a blacklist (which only blocks known bad sources).
HITL grading strategies commonly used in industry at present include: whitelists/blacklists based on action type, labels based on resource sensitivity, and dynamic triggering based on confidence thresholds.
The whitelist approach, which only declares permitted component types and properties, is safer than a blacklist because the model's output space is nearly infinite and attackers can always find variants not on the blacklist.
Whitelist/blacklist rules should be extracted into a separate configuration file (such as YAML or TOML) rather than hardcoded in the source code.
The common drawback of sandbox and whitelist mechanisms is their relatively high configuration cost and poor out-of-the-box experience, which causes many users to disable them, rendering the protections ineffective.
Whitelist (exclusion list) mechanisms and double-confirmation dialogs are common safety design features in similar process management tools.
All Facts (8)
Because models cannot fundamentally distinguish "data" from "instructions", prompt-based protection alone is unreliable; a viable approach is to have the model only output suggestions, with all actions that have side effects executed by application code according to a whitelist, and to perform schema validation and permission checks on model output.
50%UnverifiedA single-sender inbox essentially turns the entire inbox into a whitelist: by default it rejects all unknown sources, and only explicitly added senders can deliver.
50%UnverifiedA whitelist is a filtering mechanism that only allows pre-approved sources to pass, as opposed to a blacklist (which only blocks known bad sources).
50%UnverifiedHITL grading strategies commonly used in industry at present include: whitelists/blacklists based on action type, labels based on resource sensitivity, and dynamic triggering based on confidence thresholds.
50%UnverifiedThe whitelist approach, which only declares permitted component types and properties, is safer than a blacklist because the model's output space is nearly infinite and attackers can always find variants not on the blacklist.
50%UnverifiedWhitelist/blacklist rules should be extracted into a separate configuration file (such as YAML or TOML) rather than hardcoded in the source code.
50%UnverifiedThe common drawback of sandbox and whitelist mechanisms is their relatively high configuration cost and poor out-of-the-box experience, which causes many users to disable them, rendering the protections ineffective.
50%UnverifiedWhitelist (exclusion list) mechanisms and double-confirmation dialogs are common safety design features in similar process management tools.
50%