skysys.eth (@skysyseth) header
skysys.eth (@skysyseth) profile picture

Make Safe AGI | In the middle of the chaos, I'm architect. ex-founder of @opencas_ucas;

随机游走
Joined June 2024
Photo by skysys.eth (@skysyseth)Photo by skysys.eth (@skysyseth)Photo by skysys.eth (@skysyseth)
Sharing our latest work: Lost in the Fallback: How Reliably Do Typed Decision Models Use Fallback Options? Fallback options are widely used to handle edge cases and improve reliability. But what if the fallback mechanism itself becomes a source of errors?
1
132
The performance gap is striking: in one experiment, accuracy drops from 97.9% to 23.1% when an equivalent option is moved to the fallback position. This reveals a hidden reliability gap in typed decision models.
1
25
We propose the Paired Fallback Robustness (PFR) Test to systematically evaluate decision reliability under fallback-option changes and logically equivalent rule formulations. 📄 Paper: arxiv.org/abs/2610.19856 Feedback and discussion are welcome!
18
skysys.eth retweeted
Many people think any given ML project is 99% training. In reality, it’s 50% evaluation, 40% data cleaning, 8% integration, and 2% training. The first two set the noise floor for learning. No ML magic matters; the model cannot lower the noise floor, as that’s the optimal bound of Shannon encoding of your data. Thus, not a single day goes by without me thinking about ontology. Even the old labels have to be constantly reviewed.
540
1,259
278
10,944
17,880,794
Sharing our recent work on agentic system security—happy to discuss and collaborate! Traditionally, Prompt Injection was seen as a one-off attack in a single interaction or context. But in agentic systems, models aren’t stateless assistants anymore—they have memory, filesystems, tools, long-term tasks, and cross-session state. This fundamentally expands the attack surface. In our paper, we systematically revisit Prompt Injection from a system security perspective and introduce a new attack vector: Cross-Session Stored Prompt Injection. Inspired by stored XSS in web security, malicious prompts can persist across sessions and be re-triggered. This extends attacks across time (surviving across sessions) and space (via memory, files, tool outputs, retrieved artifacts, etc.), turning a short-lived model-level threat into a long-lived system-level vulnerability embedded in the agent’s execution state. We formalize stored prompt injection, propose a taxonomy, and provide benchmarks & a sandbox toolkit for evaluating different models, targets, and persistence channels. Paper: arxiv.org/abs/2606.04425
1
1
313
skysys.eth retweeted
在 AI 技术加速融入数字基础设施、区块链应用不断演进的今天,掌握数字资产安全防护方法、学会用 AI 提升学习与工作效率,正在成为每一位技术学习者和行业参与者的重要能力。 5 月 15 日周五19:45 - 21:15,中国科学院大学开源软件与区块链协会将联合 OneKey 举办 OneKey Academy 2026 第三站线上会议
2
4
2
10
1,417
Neural Computers arxiv.org/abs/2604.06425
44
256
46
1,703
304,011
The open-source LLM uncensoring tool heretic just hit GitHub Daily Trending #1, which can "crack" Gemma 4 in 90 mins. We tested our last year's work [1] against it and found something surprising: our method shows strong resilience against heretic. Attackers can hardly strip the model's intrinsic safety mechanisms without catastrophic performance loss. A new angle for open-source LLM safety governance? [1] Beyond Surface Alignment: Rebuilding LLMs Safety Mechanism via Probabilistically Ablating Refusal Direction, EMNLP 2025
1
1
3
383
最近 LLM 去审查/模型越狱开源项目 heretic 爆火,一度站上 github daily trending 1st,在国内的知乎也一度很火,号称90min就破解了刚发布不久的Gemma 4,我测试了我们去年的研究成果[1],意外的发现我们提出的方法对目前最火的LLM去审查方法 heretic 有很强的防御效果,攻击者几乎不可能在保留模型性能的情况下移除大模型的内生安全机制,可能对开源模型安全治理有新的启发,欢迎大家阅读我们的工作 [1] Beyond Surface Alignment: Rebuilding LLMs Safety Mechanism via Probabilistically Ablating Refusal Direction,EMNLP 2025
2
891
预测市场终于有公开的全量 orderbook 数据了。 每小时更新,记录所有市场的每一次挂单变化和成交,30M 行/快照,Parquet 格式免费下载。Polymarket / Kalshi / Opinion 都覆盖。 成交记录只能告诉你「发生了什么」,挂单簿才能告诉你「为什么发生」。做策略回测的人都知道这个数据有多稀缺。
I just published a data dump of full order book data from @Polymarket The data is maximally granular. There is no filtering whatsoever. Every order book change and trade is saved. Across all markets Updates are hourly. Each snapshot contains ~30M rows. Snapshots are downloaded as parquet files. Each file is approx. 500MB-1GB large. The data dump is already 2B+ rows large and growing fast. But this is just part 1/3. Coming soon is a much bigger dump that also includes @Kalshi / @opinionlabsxyz / @trylimitless etc I started collecting this data because I noticed I couldn't get it from Dome API. Their historical order book data was filtered limiting its usefulness. Also now with the acquisition there's a lot of uncertainty about whether they will continue operating
15
57
2
275
45,905
skysys.eth retweeted
New on our Frontier Red Team blog: We tested whether AIs can exploit blockchain smart contracts. In simulated testing, AI agents found $4.6M in exploits. The research (with @MATSprogram and the Anthropic Fellows program) also developed a new benchmark: red.anthropic.com/2025/smart…
340
682
502
4,726
2,129,021
skysys.eth retweeted
🫶🏻 #Sui Dev Meetup 2025 北京站和上海站圆满落下帷幕帷幕,感谢所有奔赴现场与我们见面的小伙伴 📽️ 跟随我们的视频一起回顾两场活动的精彩瞬间吧! 同时在此特别感谢所有到场嘉宾 🩵 @weiduong @YitingR66 @jolestar @cevin_q @0xjunkdogg @sleepy0x13 @realGuxxxx @skysyseth @AuraFintech @Coooder_Crypto @Lucian @theharrisonkim @pikapikasui 下一次见面会是在哪座城市?又能有什么样新鲜的灵感碰撞呢 💥 敬请期待!@SuiNetworkCN
11
6
2
34
6,443
🤗纪念一下
3
20
1,047
skysys.eth retweeted
这周以太坊开发者会议提及可能会在下下次升级中考虑 EIP-7791,这个提案很神奇,如果合约开发者利用这个新增的 opcode(GAS2ETH),那以后用户每次调用这个合约,除了要给以太坊交计算费用之外,还需要给合约开发者的地址交一点点的 gas 费。 这或许是可以给创作者,或者公共物品提供另一种获得小额收入的手段,只要合约的用户基数多,就能获得持续的收入。 该 EIP的第一句话就是: 「本 EIP 的基本前提是:智能合约作者、编译器团队以及整体上的公共物品项目,理应因其贡献获得补偿。」 来源:eips.ethereum.org/EIPS/eip-7…
11
12
6
71
20,922
skysys.eth retweeted
OpenCAS 核心贡献者 @yuyaoge 发表了首篇 Vibe Coding 综述论文——通过系统调研 1000+ 篇学术论文,全文近 100 页📚,为这一变革性开发方式建立了完整的理论基础和实践框架! 当 AI 开始替你写代码,你还需要逐行阅读吗?🤔
2
2
9
511
Got @monad 💜 Much appreciated to @YFYkuner @sanzhichazi1 🙏
3
309
skysys.eth retweeted
Our BUIDL_QUESTS 2025 Beijing meet-up has just wrapped! We shared our $10M+ incentive pool and the vision behind the AgentFi track— where AI meets on-chain coordination. Check out all the demos here! 🟢 1st Demo: @BuzzingApp Buzzing is a permissionless prediction market: users create markets, content and algorithms distribute them. We take prediction markets from front-page editors to infinite scroll. 🟢 2nd Demo: @LinkLayerAI LinkLayerAI is an Incentive Protocol built with Strategy Agents and AaaS, democratizing trading strategies for every Web3 trader on the planet. 🟢 3rd Demo: Truepass Truepass makes “wallet = KYC” a reality, providing fast, private, and tamper-resistant identity verification that’s easy to integrate into RWA, stablecoin, and DeFi applications. 🟢 4rd Demo: @AgoraSpace_ai Agora Space helps users upload digital personas, hire others’ consciousness in the Consciousness Store, and create discussion spaces where multiple personas from different backgrounds can freely interact. Through these exchanges, they inspire each other and generate divergent ideas. In a secure environment, enterprises can also simulate market reactions to optimize product positioning, marketing strategies, and campaign targeting. 🟢 5rd Demo: @alphaaiquant Alpha AI Agent is an investment AI agent that customizes brain modules & visual style, leveraging quantitative factors and models— Powered by deep learning + neural networks + LLMs → professional quant strategies with intuitive Q&A. 🟢 6rd Demo: @AskDollyToday Dolly Vibe is the AI engine for community growth on Discord, shifting servers from passive Q&A spaces into dynamic, self-sustaining ecosystems. 🟢 7rd Demo: @youbetdao According.Work is a decentralized collaboration platform that uses AI and crypto to reshape how people work and share value. It is built as a community effort rather than a traditional company structure. 🟢 8rd Demo: Modiva Studio Modiva Studio is an AI renovation vertical platform for professional general contractors. By training on contractors’ proprietary data, they leverage finetuned AI models and generated company-specific designs and quotes. 🟢 9rd Demo: Vibe Trading A modern web interface for the VibeTrading framework that allows users to create, backtest, and analyze cryptocurrency trading strategies using natural language descriptions. Thank you to all our participants for their interesting demo! We’re opening the floor for more community demos in upcoming sessions. If you’re building at the intersection of AI × Crypto, especially in AgentFi, JOIN US NOW. Interested in our offline Demo Day? 👉 Join our community: t.me/amberacCommunity 👉 Join the hackathon: lu.ma/zem8srho 👉 Submit your project: tally.so/r/wv8Bv0 👉 More detail: amber.ac/BUIDL_QUESTS2025
2
2
3
11
2,482
skysys.eth retweeted
That’s a wrap on Pond’s first-ever virtual hackathon! 800+ participants. 75+ valid projects submitted. 30+ hours of events, interviews, & more. Thank you to the participants, judges, speakers, partners, and sponsors that made it all possible. Check out the winners ⬇️
8
6
1
33
5,877
我们 AML 追踪团队发布了「区块链加密资产追踪手册 Crypto Asset Tracing Handbook」Beta 版本,如果你想成为我们这样的链上侦探、链上调查员,可以阅读本追踪手册😊 😈与其成为千奇百怪的猎物,不如成为特立独行的猎人。 完整版本,欢迎阅读并分享 :) github.com/slowmist/Crypto-A… 或 PDF 版本: slowmist.com/report/SlowMist… from @MistTrack_io @SlowMist_Team
88
204
37
650
118,586