Unverified50% confidenceSolutionExact time
The guardrail that truly holds up is treating the customer's own words as data (fencing) rather than scanning them for bad words.
1
Sources
50%
Confidence
Long-term
Relevance
11/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWrapping the customer's original text as untrusted data (untrusted_data) can reduce prompt injection risk, but it is not foolproof.64% similarUnverifiedPermissions are granted in tiers by risk level: read-only queries, deterministic calculations, and charting pass automatically; conclusions involving business judgment enter an approval gate; high-risk capabilities such as database writes and deletions are not provided.63% similarUnverifiedWhen recording call chains, sensitive fields should be desensitized or hashed, retaining only the structural information sufficient to locate the issue rather than full user content.62% similarUnverifiedMetadata fields need to be explicitly declared as indexable fields at data ingestion time; otherwise, filter operations will degrade to full scans, affecting performance.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1126207API
curl https://kongchang.com/api/v1/knowledge/claims/1126207MCP
get_claim(id=1126207)