以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。
数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15
Anthropic Mythos
34 atoms · 跨 14 天 · 首见 2026-04-07 · 最近 2026-06-23
三色: 🟦 fact 20 · 🟥 take 14 · stance ▲8/▼3/◆8
叙事状态: 🍂 fading (退潮) · 💤 沉寂 · 策展近14d 0 atom
状态只算低频策展源 (群/卖方); X firehose 仅作背景音量
来源: 群 2 · X 32
时态: fresh:33 · aging:1
标签: 好数字:17 · 好信源:10 · 好观点:10 · 好思考:5 · 好问题:1
🟨 AI 综合 · junior analyst 概览
展开 AI 综合 (灰色 · 非市场结论 · 点击数字溯源到原 atom)
model: deepseek-chat · 2026-07-12 · 默认折叠
Anthropic Mythos 是一款专注于自主发现和连锁漏洞的受限前沿安全模型
¹。Palo Alto Networks 使用 Mythos 在三周内发现了超过两打关键漏洞，且消耗了超过 100 万美元的代币
¹¹¹。该模型能够识别高级人类专家和现有 AI 系统遗漏的网络漏洞
¹，并已在美国系统中发现漏洞
¹。定价方面，Mythos 的定价是 Opus 的 5 倍
¹。看多者认为 Mythos 是优于 OpenAI 任何模型的超级模型
¹，若向所有付费用户开放，计算需求将一夜增加 10-15 倍
¹。但看空者指出，仅使用该模型数周就可能花费数百万美元
¹，且预期其活跃参数比例可能低至 2% 而非 5%
¹。若 Mythos 无法解决 5.5 版本能解决的问题，将证明 Anthropic 为维持 SWE 优势而过度专业化
¹。
🧭 拥挤度 (一人一票): ▲ 2 位作者 (KOL2) + 群 1 条 vs ▼ 3 位作者 (KOL3)
🔗 因果传导 (causal map)
*Anthropic Mythos 在产业链上的传导关系. 边是群里/卖方陈述的因果 (非 AI 推断), 数字 = 几条 atom 支撑. 点 atom 溯源.*
↓ 下游·近期陈述 (1)
- 🟢 AI 算力产业 · 需求拉动 · 1 次陈述
如果向所有付费用户开放，将在一夜之间增加 10-15 倍的计算需求
🔮 前瞻触发器 (forward triggers)
*这些 atom 指向可能 reprice 的前瞻事件 (无精确日期, 仅类型). 配合上面叙事状态看「上膛」程度.*
产品 (1 · ▲1/▼0)
- 🟢 2026-04-07 · 全量开放后的计算需求增量
如果向所有付费用户开放，将在一夜之间增加 10-15 倍的计算需求
🟦 客观事实 (facts) (19)
- <a id="atom-285d9bfacc8a6374"></a>🟦 2026-06-01
X·@theinformation ⭐ · used by Palo Alto Networks to find critical vulnerabilities · 产品发布/合作伙伴
Palo Alto Networks used Anthropic’s Mythos to find more than two dozen critical vulnerabilities in about three weeks.
tags: 好数字·好信源 · → daily
value: qty=more than two dozen
- <a id="atom-2b1096748e002c59"></a>🟦 2026-06-02
X·@theinformation · 发现 Palo Alto Networks 关键漏洞数量 · 安全漏洞/合作客户
Palo Alto Networks says Anthropic’s Mythos found more than two dozen critical vulnerabilities in about three weeks—and consumed more than $1 million in tokens.
tags: 好数字·好信源 · → daily
value: qty=more than two dozen · date=about three weeks
- <a id="atom-68b967c0f3dc33de"></a>🟦 2026-06-02
X·@theinformation · Token 消耗金额 · 资本开支/技术路线
Palo Alto Networks says Anthropic’s Mythos found more than two dozen critical vulnerabilities in about three weeks—and consumed more than $1 million in tokens.
tags: 好数字·好信源 · → daily
value: qty=$1 million
- <a id="atom-d40af090aa4473f2"></a>🟦 2026-04-07
群 · 定价 vs Opus · 定价比较
Anthropic 发布的 Mythos 模型定价是 Opus 的 5 倍
tags: 好数字 · → daily
value: qty=5x
- <a id="atom-2fdf03bee794b719"></a>🟦 2026-06-14
X·@DrNHJ · 识别了高级人类专家和现有AI系统遗漏的网络漏洞 · 网络安全
Mythos가 고급 인간 전문가와 기존 AI 시스템도 놓친 사이버 취약점을 식별할 수 있다는 점
tags: 好数字 · → daily
- <a id="atom-f7c580d27487038a"></a>🟦 2026-06-23
X·@zerohedge · found vulnerabilities in US systems · 安全事故
ANTHROPIC’S MYTHOS FOUND VULNERABILITIES IN US SYSTEMS: AP
tags: 好信源 · → daily
- <a id="atom-0f08c1ea9bef34a8"></a>🟦 2026-06-21
X·@dreamworks2050 · is a restricted frontier AI model specialized in cybersecurity tasks like autonomously discovering and chaining vulnerabilities · 技术路线
Mythos is Anthropic’s restricted frontier AI model specialized in cybersecurity tasks like autonomously discovering and chaining vulnerabilities.
tags: 好信源 · → daily
- <a id="atom-0354e4dc45f41a6f"></a>🟦 2026-06-04
X·@scaling01 · 训练适用C=6ND范围 · 算力估算
for all parts of training that can be approximated by C=6ND so RL for example is not included here
tags: 好信源 · → daily
- <a id="atom-4cbf83c54633575e"></a>🟦 2026-05-09
X·@dreamworks2050 · bug fixes shipped · 产品发布
April 2026: 423 bug fixes shipped vs just 31 exactly a year earlier
tags: 好数字 · → daily
value: qty=423 · date=April 2026
- <a id="atom-1af54ab3dd6acf3d"></a>🟦 2026-05-09
X·@dreamworks2050 · found dormant bugs · 产品发布
Found bugs dormant for 15+ years in the codebase
tags: 好数字 · → daily
value: qty=15+ years · date=2026-05-09
- <a id="atom-59da6d0223e2d250"></a>🟦 2026-05-09
X·@dreamworks2050 · discovered sandbox vulnerabilities · 产品发布
Discovered sandbox vulnerabilities (normally $20K bounty each) at higher volume than human researchers ever did
tags: 好数字 · → daily
value: qty=higher volume than human researchers · date=2026-05-09
- <a id="atom-c280524085f8bfec"></a>🟦 2026-05-08
X·@chatgpt21 · METR 50% time horizon · 产品发布
Mythos has a 50% time horizon of at least 16 hours.
tags: 好数字 · → daily
value: qty=16 hours · date=2026
📷 原图
- <a id="atom-c7f538991a2d0371"></a>🟦 2026-05-08
X·@chatgpt21 · public time horizon evals rank · 评测结果
Mythos at #1 beyond a shadow of a doubt on public time horizon evals.
tags: 好数字 · → daily
value: direction=#1
📷 原图
- <a id="atom-f32d98aad5eca353"></a>🟦 2026-04-30
X·@Techmeme [老化中] · 被 NSA 用于寻找软件漏洞 · 合作客户
Sources: the NSA has been testing Anthropic's Mythos model to find vulnerabilities in Microsoft products and widely used software from other companies
tags: 好信源 · → daily
- <a id="atom-aaf48e59a6724b71"></a>🟦 2026-06-20
X·@pmddomingos · broke into almost all of the NSA’s classified systems in hours · 安全事故
Mythos broke into almost all of the NSA’s classified systems in hours, per its director.
tags: 好信源 · → daily
value: direction=hours
- <a id="atom-96dc534b565a4542"></a>🟦 2026-06-14
X·@zerohedge · 被美国出口禁令封锁外国接入 · 出口管制/地缘政治
Canada Prime Minister Mark Carney said the US export ban blocking foreign access to Anthropic's Mythos underscores the risk of depending on just a handful of powerful AI tools.
tags: 好观点 · → daily
- <a id="atom-192893b16f58f227"></a>🟦 2026-06-04
X·@javi_22_dev · 训练token量 · 训练数据量
200-300T training tokens? Total seen or pretraining only?
tags: 好问题 · → daily
value: qty=200-300T · date=2026
- <a id="atom-dfa9291a5c84521a"></a>🟦 2026-05-09
X·@dreamworks2050 · compromised patch capability · 技术路线
Mythos writes a compromised patch, then attacks the browser with it — a multi-step creative process
tags: 好观点 · → daily
- <a id="atom-cdd298d5de0fc1b2"></a>🟦 2026-05-09
X·@dreamworks2050 · AI-generated patches deployment status · 产品发布
AI-generated patches still can't be deployed directly. Every fix is "one engineer writing a patch, one engineer reviewing it." Not yet automatable.
tags: 好观点 · → daily
🟥 多头 takes (bullish) (4)
- <a id="atom-00c1ebd5b4546525"></a>🟥 2026-06-04
X·@scaling01 ⭐ · 活跃参数比例原因 · 模型稀疏性
Recent papers also put the optimal total/active ratio at 18-25, while other papers mention that reasoning ability peaks around 20 tokens per parameter. I think 5% sparsity is reasonable.
tags: 好信源·好思考 · → daily
- <a id="atom-5cf68c34a35e8cb1"></a>🟥 2026-04-07
群 · 全量开放后的计算需求增量 · 算力需求增长/全量开放
如果向所有付费用户开放，将在一夜之间增加 10-15 倍的计算需求
tags: 好数字 · → daily
value: qty=10-15 倍
- <a id="atom-e351a3f7bf255fd5"></a>🟥 2026-06-04
X·@lukeburgis · 模型质量超越OpenAI · 技术路线/模型发布
Anthropic Mythos is a superior model to anything OpenAI has built
tags: 好观点 · → daily
- <a id="atom-e546a2710482effd"></a>🟥 2026-06-04
X·@lukeburgis · 更强安全关注度 · 技术路线
they simultaneously pay more attention to safety and security
tags: 好观点 · → daily
🟥 空头 takes (bearish) (3)
- <a id="atom-97a7175982027c28"></a>🟥 2026-06-01
X·@pstAsiatech · cost for weeks of use · 资本开支/盈利能力
using the model even for just a couple weeks can cost millions of dollars
tags: 好数字·好观点 · → daily
value: qty=millions of dollars · date=couple weeks
- <a id="atom-f01456c0f240f340"></a>🟥 2026-06-04
X·@jakehalloran1 · 活跃参数比例预期 · 模型稀疏性
I expect it to be less active than that imo Like 2% not 5
tags: 好观点 · → daily
value: qty=2% · date=2026
- <a id="atom-c43c30c1ae6c98c2"></a>🟥 2026-05-27
X·@teortaxesTex · 若无法解决同样问题则证明过度专业化以维持SWE优势 · 竞争格局
If mythos could not solve the same problem that 5.5 solves, it'd be literally over for Anthropic, a proof that they had to overspecialize a 10T class model to maintain an edge in SWE
tags: 好思考 · → daily
🟥 中性 takes (neutral) (4)
- <a id="atom-d635c02befaf4a40"></a>🟥 2026-06-04
X·@scaling01 ⭐ · 训练计算成本估计 · 算力估算
my new median is now a bit higher at 12.98T @ 649B
tags: 好数字·好思考 · → daily
value: qty=12.98T · date=2026
📷 原图
- <a id="atom-cca705186e91c720"></a>🟥 2026-06-04
X·@scaling01 ⭐ · 训练规模相对Opus倍数 · 成本对比
Mythos being 5x more expensive than Opus, for which my estimate was 2-3T
tags: 好数字·好思考 · → daily
value: qty=5x · date=2026
📷 原图
- <a id="atom-1bd618574d54d526"></a>🟥 2026-06-04
X·@scaling01 · 训练时长下限 · 训练时长
use a minimum of 2 months on 500k Trainium 2 chips
tags: 好思考 · → daily
value: qty=2 months · date=2026
📷 原图
- <a id="atom-09ca7633b45a1553"></a>🟥 2026-06-08
X·@zekramu · compared to Opus/5.5 xhigh at security tasks · 模型能力/竞争格局
far better than opus / 5.5 xhigh
tags: 好观点 · → daily
🟦 新闻流 (squawk · 1)
展开新闻流 (FirstSquawk / financialjuice / wallstengine / DeItaone — 快讯, 非原创 take)
- <a id="atom-60b4d669035127a5"></a>🟦 2026-06-18
X·@financialjuice · 早期用户保留访问权限（美国指令后） · 产品发布/监管政策
Early users of Anthropic Mythos retain access following US directive
tags: 好信源 · → daily
⏱ 时间轴 (近 20)
- 🟦 2026-06-23 ·
fact · found vulnerabilities in US systems · → daily
- 🟦 2026-06-21 ·
fact · is a restricted frontier AI model specialized in cybersecuri · → daily
- 🟦 2026-06-20 ·
narrative · broke into almost all of the NSA’s classified systems in hou · → daily
- 🟦 2026-06-18 ·
fact · 早期用户保留访问权限（美国指令后） · → daily
- 🟦 2026-06-14 ·
narrative · 被美国出口禁令封锁外国接入 · → daily
- 🟦 2026-06-14 ·
fact · 识别了高级人类专家和现有AI系统遗漏的网络漏洞 · → daily
- 🟥 2026-06-08 ·
fact · finds ~800 major vulnerabilities across products · → daily
- 🟥 2026-06-08 ·
narrative · compared to Opus/5.5 xhigh at security tasks · → daily
- 🟥 2026-06-04 ·
forecast · 训练时长下限 · → daily
- 🟥 2026-06-04 ·
forecast · 训练计算成本估计 · → daily
- 🟥 2026-06-04 ·
forecast · 训练规模相对Opus倍数 · → daily
- 🟦 2026-06-04 ·
forecast · 训练token量 · → daily
- 🟥 2026-06-04 ·
forecast · 活跃参数比例预期 · → daily
- 🟦 2026-06-04 ·
fact · 训练适用C=6ND范围 · → daily
- 🟥 2026-06-04 ·
narrative · 活跃参数比例原因 · → daily
- 🟥 2026-06-04 ·
narrative · 模型质量超越OpenAI · → daily
- 🟥 2026-06-04 ·
narrative · 更强安全关注度 · → daily
- 🟥 2026-06-04 ·
forecast · RL与基模型计算比 · → daily
- 🟥 2026-06-04 ·
forecast · RL占比展望 · → daily
- 🟦 2026-06-02 ·
fact · 发现 Palo Alto Networks 关键漏洞数量 · → daily
← 实体目录 · 系统日志
