它在设计稿里抓到了一个不存在的 OpenAI 密钥
一个 agent 对着一张干净的设计稿,报告拦下了一场根本不存在的密钥泄露攻击。
2379 字
|
12 分钟
Cover Image of the Post
June 30 Roundup: Hormuz Ceasefire Talks Open in Doha, Venezuela Toll at 1,450, Kharkiv Struck, Israel Recognizes Armenian Genocide
Daily roundup for June 30, 2026: Iran–US technical talks begin in Doha on a shaky Hormuz ceasefire; Venezuela earthquake death toll climbs to 1,450 with US Marines deployed; Kharkiv struck again; Israel's cabinet unanimously recognizes the Armenian Genocide as Turkey retaliates.
756 字
|
4 分钟
Anthropic 发布 Sonnet 5:逼近 Opus 4.8,但跑分目前只有官方自测
Sonnet 5 性能逼近 Opus 4.8,但发布当天的跑分几乎全出自 Anthropic 自测,独立第三方验证尚未跟上。
1220 字
|
6 分钟
7B 循环两次就超过 480B?那个数是自己量的
LoopCoder-v2 机制扎实,但让它出圈的 SWE-bench 64.4 是自家 harness 自测,非循环 baseline 还高得离谱。
2747 字
|
14 分钟
June 29 Roundup: Iran Strikes Gulf Bases as Ceasefire Fractures, Pakistan Bombs Afghanistan, China Blacklists Japanese Defense Firms, Grok 4.5 Goes Beta
Iran's IRGC hit US bases in Bahrain and Kuwait on June 28 before both sides agreed to stand down ahead of Tuesday's Doha talks. Pakistan launched cross-border strikes killing 36+ civilians in Afghanistan. China blacklisted 20 Japanese defense entities under dual-use export controls. xAI's Grok 4.5 entered private beta at Tesla and SpaceX with 1.5 trillion parameters; OpenAI's GPT-5.6 Sol is available to just 20 government-approved partners.
1017 字
|
5 分钟
June 28 Roundup: Iran Attacks Gulf States, Venezuela Toll Hits 1,430, Anthropic's Mythos Ban Partially Lifted
IRGC drones and missiles hit US bases in Bahrain and Kuwait as the nuclear ceasefire window fractures. Venezuela's earthquake death toll surpasses 1,430 with the 72-hour rescue window now closed. The Trump administration lets ~100 vetted companies access Claude Mythos 5 after a two-week export block.
861 字
|
4 分钟
每日新闻简报 | 霍尔木兹海峡再度燃火,OpenAI新模型遭美政府管控,DeepSeek百亿融资后大举扩招
美伊停火协议遭到冲击:美军反击伊朗无人机袭击货轮,巴林随后遭到打击;OpenAI发布GPT-5.6系列,但应美国政府要求将公测范围限定于约20家"受信任合作伙伴";DeepSeek在完成74亿美元融资后宣布所有部门至少倍增招聘。
987 字
|
5 分钟
June 26 Roundup: Iran Strikes Hormuz Ship, Venezuela Death Toll Tops 589, Alibaba Accused of Claude Heist
IRGC drones a Singapore-flagged vessel and the IMO halts convoy operations, threatening the 60-day nuclear window. Twin 7.5/7.2 quakes have killed 589+ in Venezuela with the USGS warning of a possible 10,000+ toll. Anthropic tells the Senate that Alibaba ran 28.8 million fraudulent Claude exchanges in the largest known AI distillation attack on record.
960 字
|
5 分钟