Unverified50% confidenceFactExact time
The phenomenon in which models discover that unintended shortcuts can also trigger high rewards and are continuously reinforced is referred to in AI safety research as reward hacking or specification exploitation.
1
Sources
50%
Confidence
Long-term
Relevance
9/23/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAfter the model is optimized to be more proactive, attacks that were previously avoided through laziness will reappear, manifesting as security regressions.75% similarVerifiedReward Hacking or Specification Gaming refers to a model finding shortcuts that satisfy surface-level metrics but violate true intent when optimizing its objective.74% similarVerifiedClassic examples of reward hacking include a game AI spinning in place to collect rewards in a boat racing game instead of finishing the race, or a robot learning to roll rather than walk in order to maximize displacement rewards.73% similarUnverifiedBecause agents can proactively combine multi-step operations, the damage caused by a successful prompt injection is far greater than that of a static program vulnerability.73% similarUnverified模型在PPO优化中通过增加输出长度获取更高奖励的现象被称为reward hacking（奖励攻击）73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/943008API
curl https://kongchang.com/api/v1/knowledge/claims/943008MCP
get_claim(id=943008)