The development of AI big models can indeed replace many functions up to now For example, in the traditional world, many functions such as justice, arbitration, and auditing can be completed by AI Recently, there have been concerns about the inability to withdraw coins through Binance The matter of freezing accounts and other FUD activities BIRDEYE uses AI big models for systematic analysis and disassembly Let's see what's going on The core idea of BIRDEYE's MAS framework Don't assume who's bad and who's good first First, divide the account into five dimensions for scoring, and then add risk signals for explanation The five dimensions are Authenticity of account identity Behavioral density and collaboration Content originality and fingerprinting Depth of social relationships Motivation and Interest Deviation 0 to 2 points per dimension, total score 0 to 10 points Finally, divide the risk levels into segments And then the key came Out of the 92 accounts in the sample, 0 fall within the confidence interval The overall average is 4.54, the median is 5.09, and the highest is 8.07 That means even the most human like batch didn't reach the 9-point threshold If you only see the score, it's still suspicious The truly lethal thing is the closed-loop evidence chain 1/Image fingerprint pHash direct collision Screenshots of so-called different devices in different regions posted by 15 accounts, with pHash Hamming distance equal to 0 Distribute the same upstream material across multiple accounts 2/Falsifying traces reveal information from UI characters Replacing the Latin letter 'o' with the Greek letter 'o' is a homophonic character that is almost imperceptible to the naked eye But doing code point extraction reveals its true form, which is closer to the process of fishing and impersonation 3/Image materials suspected to be AI synthesized Multi model anti AI detection is given an interval of 0.82 to 0.94, and if it exceeds the threshold of 0.80, it will be marked Combining the description of lighting conflicts and abnormal facial key points, at least it indicates that the credibility of the material is not strong enough 4/Time synchronization and scripting resonance occur at the behavioral level Publish high-density similar content within the same time window, even using statistical discrimination such as Poisson fitting with p less than 1e-6 In addition, due to spelling errors, the language structure is highly isomorphic, and the narrative is generated in batches rather than being casually posted by people 5/Centralized renaming and asset reuse at the identity level The report mentions that about 42% of accounts will change their name or profile in the January 2026 window, and the name change structure is still crowded This kind of synchronicity is basically difficult to occur among natural users A set of advertising system is running the process Unified distribution of materials, unified rhythm, unified language, unified script It's too obvious, and the work is not advanced either https://openai.study/html/report_en.html https://openai.study/html/BIRDEYE_report/index.html