Unverified50% confidenceSolutionExact time
Alignment reduces the probability of deviation, while constraints limit the consequences of deviation; together they form defense in depth.
1
Sources
50%
Confidence
Long-term
Relevance
11/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedPermission boundaries can prevent a model deceived by prompt injection from turning intent into actual damage, and are a relatively low-cost and reliably effective part of defense in depth.77% similarUnverifiedSoft constraints encourage models to reduce violations by adding penalty terms to the loss function, but can only lower the probability of violations rather than fundamentally eliminating them.77% similarUnverifiedSilent patches are intended to reduce risk, but in practice may expand the attack surface by delaying users' proactive defenses.74% similarUnverifiedNegative constraints narrow the probability space by explicitly excluding undesired outputs, are equally important as positive instructions, and can reduce deviations caused by the model's creative improvisation.74% similarUnverifiedThe fidelity improvement brought by the risk-seeking method does not come without a cost: average fidelity decreases accordingly.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1169292API
curl https://kongchang.com/api/v1/knowledge/claims/1169292MCP
get_claim(id=1169292)