| anonymous caller of a public endpoint | owner's money | per-caller allowance; shared limit with holds; free-only fallback |
|---|---|---|
| a caller's own key being misused by a loop or bug | the person's money | optional caller-chosen cap; limit on the key at the provider |
| a misbehaving or hostile model | the application, through its output | validation of structured answers; answer treated as untrusted input |
| a provider that keeps or trains on prompts | private text | private mode; refusal of providers that publish nothing |
| a writer to the shared store | availability and cost | shape and size checks on everything read back |
| compromise of the package or its release | every downstream user | no dependencies; pinned build; provenance; protected release environment |
| leakage through logs | prompts, keys | no logging of prompts or answers; keys never stored; error bodies not enumerable |
AI & evaluation · Updated 2026-11
Adversaries and faults, the asset at risk and the main mitigation
Table 2 in “Lodestone: Keeping AI Calls Working on Unreliable Free Models Inside a Serverless Time Limit”: 7 rows, 3 columns.
Cite this chart
Extrafemi Research (2026). Adversaries and faults, the asset at risk and the main mitigation [Chart]. Data from Lodestone: Keeping AI Calls Working on Unreliable Free Models Inside a Serverless Time Limit. https://research.extrafemi.com/data/lodestone--data-2