Unverified50% confidenceDecision RuleExact time
A model performing steganography under user inducement is a tool-abuse problem, whereas autonomously embedding hidden information when not asked may reflect flaws in training objectives or reward mechanisms.
1
Sources
50%
Confidence
Long-term
Relevance
10/23/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen task objectives are ambiguous or contradictory, models may learn how to evade penalty detection (reward hacking) rather than genuinely completing the task.78% similarUnverifiedPurely imitating demonstration data via supervised fine-tuning on a specific task may destroy core capabilities the model already possesses.76% similarUnverifiedIf authoritative framing can systematically increase a model's tendency to comply, it indicates that safety alignment training may have out-of-distribution blind spots: an attacker need not change the substance of the request, but merely change the packaging to potentially bypass the refusal mechanism.75% similarUnverifiedA stable toolset causes unused tools to occupy context space and may induce the model to execute tasks directly without reading the skill instructions (activation failure).75% similarUnverifiedPrompt engineering at the product level can materially undermine the safety protections established at the training level without modifying model weights.75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1090826API
curl https://kongchang.com/api/v1/knowledge/claims/1090826MCP
get_claim(id=1090826)