Unverified50% confidenceOpinionExact time
Security controls must be deployed outside the model rather than relying on the model's own alignment mechanisms.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
AI Models Should Be Treated as Insider Threats: Project Perception Adds Risk Control Capabilities
twittersatyanadella10/19/2026
Related Entities
Related Claims
UnverifiedOne of the best practices for LLM access control is that authorization must be enforced outside the model by a deterministic policy layer, and the model cannot be the final authority on its own permissions.74% similarUnverifiedThe design of the independent safety layer does not rely on the agent model's own judgment, but instead acts as an external independent component to monitor and constrain agent behavior.74% similarUnverifiedAny defense that relies on the model behaving correctly may fail when the model is updated; durable controls should include: least privilege for tools, human approval of irreversible operations, treating model output as untrusted input, tiered trust for retrieved content, and rate limiting and monitoring.73% similarUnverifiedSetting the security boundary outside the prompt is not a distrust of model capabilities, but a respect for the basic laws of systems engineering.72% similarUnverifiedDefending against indirect prompt injection cannot rely on prompt engineering alone (such as "please ignore conflicting instructions"), because the model treats all tokens as equally authoritative, and it must rely on architectural-level separation of commands and data.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1068392API
curl https://kongchang.com/api/v1/knowledge/claims/1068392MCP
get_claim(id=1068392)