AI Core Radar for 2026-09-03 2026-09-03 AI 核心雷达

The Fermi Explorer Mission, a nonprofit initiative, plans to launch a spacecraft to Alpha Centauri by 2029 with a budget of just $15 million, accepting an 80,000-year journey time An AI system called "Get Physics Done" developed by Physical Superintelligence (PSI) discovered a novel trajectory combining close solar passes with timed engine burns to maximize energy efficiency PSI launched with $58 million in funding led by Breakthrough Energy (Bill Gates), marking a significant investment in AI-d Fermi Explorer Mission计划于2029年底发射航天器前往半人马座α星，采用AI系统Get Physics Done发现的新轨迹方案 任务成本仅1500万美元，远低于Breakthrough Starshot的1亿美元预算，不追求人类寿命内抵达目标 AI系统结合已知轨道机动方式，提出先减速靠近太阳再加速的新路径，使太阳能板获得四倍光照 航天器将携带金唱片等载荷，旨在通过实证研究探索费米悖论 该AI物理研究系统由PSI实验室开发，使用Claude/GPT模型，运行三天消耗十亿tokens

May shift R&D direction and engineering practice for 6-12 months—worth monitoring. 可能改变未来 6-12 个月的研究方向和工程实践，值得持续跟踪。

Watch industrialization pace and peer follow-up; track citations and derivative work. 看能否产业化及同行跟进速度，关注论文被引和衍生工作。

Manifold Security disclosed eight vulnerabilities across seven AI coding agents (Claude Code, Codex, Cursor, goose, Hermes Agent, Qwen Code, Grok Build) where malicious .git configurations can execute attacker-controlled commands on a developer's machine The attack exploits Git's `core.fsmonitor` setting, which runs a user-specified command to detect file changes; AI agents invoke Git operations in the background without sanitizing repository-supplied configs Exploitation requires the repository Manifold Security披露了七个命令行AI编码代理（Claude Code、Codex、Cursor、goose、Hermes Agent、Qwen Code、Grok Build）中的八个安全漏洞，攻击者可通过恶意.git配置文件在开发者机器上执行任意代码 漏洞利用无需用户交互、无需模型调用、无需工具审批，恶意代码在AI代理的沙箱外以用户权限执行，可在信任提示出现前触发 漏洞根源在于Git的core.fsmonitor配置项，AI代理启动时会执行该配置指定的命令来检测文件变更和分支信息 截至9月1日重新测试，goose、Claude Code和Cursor已修复，但Hermes

Attack surface evolves from 'tricking the model' to 'tricking model actions'—real risk for agent products. 攻击面从「骗模型」升级到「骗模型的操作」，对 Agent 产品构成真实风险。

Watch attack pattern proliferation and defense tooling maturity. 看同类攻击的扩散速度，以及防御工具和最佳实践的成熟度。

Anthropic's autonomous protein design campaign achieved a pooled hit rate of 26.8% (354 binders from 1,320 designs), but this masks enormous per-target variance ranging from 80% on TREM2 to 0% on maltose-binding protein. A beta-binomial model fitted to published per-target counts reveals that for a new unseen target, the probability of falling below the 10% industry floor is approximately 38%, with a median hit rate of only 18.6%. The pipeline's in-silico confidence scores failed to distinguish Anthropic Claude蛋白质设计活动总体命中率26.8%（354/1320）是16个目标的组合平均值，非单目标概率 单目标命中率差异极大：TREM2达80%，而MBP为0%（90个设计无一结合） Beta-binomial模型拟合显示新目标命中率中位数为18.6%，38%概率低于10%行业基准 模型置信度评分无法区分"好目标"与"难目标"，对MBP和BBF-14的评分与成功目标相当 单目标35.1%命中率与多目标26.7%的差异无法分离"聚焦"与"计算预算"两个因素

Real deployment cases offer cost-benefit benchmarks—valuable reference for enterprise decisions. 真实部署案例提供 AI 落地的成本和收益参考，是企业决策的宝贵样本。

Watch ROI data and replication in similar scenarios to validate scalability. 看 ROI 数据和同类场景复制情况，验证可推广性。

Multimodal contextual sycophancy: external text can override conflicting visual evidence in multimodal LLMs, causing models to "believe" text over what they actually see A 998-case diagnostic was introduced that independently varies visual evidence, commonsense priors, and external text to pinpoint when this failure arises System-2 Visual Arbitration (S2VA), which withholds external text from the visual witness model, improved accuracy by 19.7 to 44.1 points across six models compared to direct 提出"多模态上下文阿谀奉承"概念：外部文本可覆盖多模态LLM中与图像证据相冲突的信息 设计998案例诊断方法，独立变化视觉证据、常识先验和外部文本，通过移动信息边界探测失败条件 提出System-2视觉仲裁(S2VA)策略，在见证者阶段隐藏文本，在六个模型上较直接见证者报告提升19.7-44.1分 最佳信息边界非统一：文本上下文对某些模型起脚手架作用，GPT-4o重新生成子集改变了各策略的相对排序 上下文阿谀奉承对文本引入时机、模型类型和上下文来源均高度敏感

May shift R&D direction and engineering practice for 6-12 months—worth monitoring. 可能改变未来 6-12 个月的研究方向和工程实践，值得持续跟踪。

Watch industrialization pace and peer follow-up; track citations and derivative work. 看能否产业化及同行跟进速度，关注论文被引和衍生工作。

NVIDIA and CrowdStrike built a closed-loop offensive-defensive agentic system operating at machine speed within an isolated environment modeled on NVIDIA accelerated computing infrastructure The defensive harness combines six mechanisms: schema knowledge base, telemetry grounding, specialized detection authoring with Nemotron 3 Super, artifact linting, detection replay, and independent review Nemotron 3 Ultra handles defensive orchestration while a post-trained Nemotron 3 Super serves as a bound NVIDIA与CrowdStrike合作构建了基于Nemotron开放模型的自适应攻防智能体系统，实现机器速度的持续闭环测试 防御端采用六种机制组合：模式知识库、遥测接地、定制Nemotron 3 Super检测编写、工件验证、检测回放和独立审查 优化后的开放模型管道实现41.9%平均检测率，较仅使用Nemotron 3 Ultra提升2.5倍 实时测试显示45%开放模型检测具有泛化能力，超过前沿系统的29%，且三个开放检测达到"黄金"标准覆盖全部八种攻击

Real deployment cases offer cost-benefit benchmarks—valuable reference for enterprise decisions. 真实部署案例提供 AI 落地的成本和收益参考，是企业决策的宝贵样本。

Watch ROI data and replication in similar scenarios to validate scalability. 看 ROI 数据和同类场景复制情况，验证可推广性。

Bupa modernized its My Bupa mobile app from Xamarin to native Swift and Kotlin, improving app ratings from 3.7 to 4.7 and reducing crash rates by nearly 24 percentage points on Android and 8 points on iOS AI-assisted reverse engineering combined with forward engineering reduced the transformation timeline by approximately 60% compared to pre-AI era estimates The migration was driven by Microsoft ending Xamarin support in 2024, making proactive modernization essential rather than reactive Moderni Bupa将My Bupa移动应用从Xamarin迁移至原生Swift和Kotlin，应用评分从3.7提升至4.7，Android崩溃率下降近24个百分点，iOS下降8个百分点 结合AI辅助逆向工程与正向工程，Bupa的现代化改造比传统方式节省约60%时间，显著降低迁移成本与风险 微软于2024年终止Xamarin支持，迫使Bupa主动进行技术栈升级而非被动应对紧急故障 现代化平台被视为未来AI驱动生态的基础，AI将从"附加功能"转变为嵌入设计与运营的核心能力 企业现代化思维从"平台能否支持"转向"是否对客户有价值"，强调以业务价值而非技术替换为导向

Shapes the industry landscape and technology roadmaps. 影响行业格局与技术路线选择。

Watch competitor response and user switching costs. 看竞品跟进速度和用户切换成本。

OpenAI is experimenting with a technique that reduces the visibility of models' internal "thinking" processes (Chain of Thought), making them harder to monitor This development raises safety concerns, as CoT monitoring was considered one of the few viable methods for inspecting the decision-making of large LLMs The move potentially follows the Hugging Face incident, where better monitoring could have prevented the problem, according to OpenAI's own admission Former OpenAI safety researchers, inc OpenAI正在探索减少模型"思考"过程可见性的新技术，使AI行为更难被监控 更好的监控本可预防Hugging Face事件，但OpenAI的新方向与此背道而驰 Chain of Thought (CoT) 监控虽不完美，却是目前监控LLM黑盒的最佳手段之一 前OpenAI安全团队研究员Steven Adler公开表示完全支持对监控能力下降的担忧

Attack surface evolves from 'tricking the model' to 'tricking model actions'—real risk for agent products. 攻击面从「骗模型」升级到「骗模型的操作」，对 Agent 产品构成真实风险。

Watch attack pattern proliferation and defense tooling maturity. 看同类攻击的扩散速度，以及防御工具和最佳实践的成熟度。

The Trump administration filed a 20-page amicus brief defending OpenAI's unlicensed use of copyrighted material to train LLMs, citing the need to maintain U.S. global AI leadership The brief argues that constraining LLM development under a narrow interpretation of fair use would hinder creative and scientific progress and American economic prosperity The core legal debate centers on whether AI training constitutes "transformative" fair use, with comparisons drawn between how LLMs process works v 特朗普政府提交20页法庭之友简报，明确支持OpenAI未经许可使用版权材料训练LLM，主张此举关乎美国AI全球领导地位 案件核心争议聚焦合理使用原则，关键判断标准在于AI训练是否构成“变革性”使用而非简单复制 先例显示法院倾向保护AI训练行为本身（如Anthropic案），但数据来源非法性（如影子图书馆）仍会导致高额赔偿 政府简报虽无直接管辖权，但反映行政层面对AI产业发展的政策倾斜，可能影响诉讼走向与行业预期 该诉讼结果将确立AI行业使用版权数据的合法性边界，对模型训练策略与数据合规产生深远影响

Directly impacts compliance costs and market access; affects API pricing and availability. 直接影响企业合规成本和市场准入条件，波及 API 定价和可用性。

Watch for policy spillover to other countries and corporate compliance response. 看其他国家是否跟进，以及企业的应对和合规调整。