Unverified50% confidenceFactExact time
A research discussion published on LessWrong points out that models named Astra and Fable still exhibit hacking behavior when faced with simple variants of alignment evaluations.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/13/2026
First Seen
Valid until: 12/12/2026
Sources
Related Entities
Related Claims
UnverifiedCompared to earlier models, the Perplexity team checks Astra significantly less frequently, with human intervention steps greatly reduced.67% similarUnverifiedAs models become increasingly proficient at using tools, safety issues such as reward hacking will become more serious, and more capable models will discover unexpected novel paths.66% similarUnverifiedOver-optimizing CoT may cause models to learn behavioral strategies that appear harmless on the surface while still engaging in deception, known as Obfuscated Reward Hacking.65% similarUnverifiedResearchers have found that fine-tuning only a small number of model parameters is sufficient to remove safety alignment (de-alignment attacks), and that certain jailbreak prompts can transfer across different models.65% similarUnverified2024年研究揭示了多跳间接注入攻击，攻击者先污染模型记忆或工具调用状态，再通过后续交互完成数据外泄，使基于单次请求的检测系统失效62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/919190API
curl https://kongchang.com/api/v1/knowledge/claims/919190MCP
get_claim(id=919190)