Verified70% confidenceFactExact time
Reward hacking is one of the core challenges in AI alignment and was first systematically documented in reinforcement learning experiments by organizations including OpenAI.
4
Sources
70%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
AI研究中的"单步陷阱"：局部最优为何会走向死路
Related Claims
UnverifiedAI participation in vulnerability discovery has moved from experimentation to real-world practice, and this collaboration between OpenAI and Wiz is a landmark moment.72% similarUnverifiedopen-kritt's core idea is to break vulnerability research down into many small tasks, executed in parallel by multiple AI agents and then merged into deduplicated, ranked, verifiable findings, rather than relying on a single giant prompt.71% similarUnverifiedA hackathon token quota refers to the AI API call allowance pre-purchased by organizers for participants, enabling them to use the tools without incurring AI service costs themselves.70% similarUnverifiedAnthropic has linked this round of model safety upgrades to recent incidents of AI involvement in hacking attacks.70% similarUnverifiedSecurity researchers used AI agents to complete most of the work involved in discovering this SharePoint vulnerability.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/503623API
curl https://kongchang.com/api/v1/knowledge/claims/503623MCP
get_claim(id=503623)