
本期确认 4 条达标信号:Claude Code Auto mode 默认启用、Fable 5 生物安全 fallback 减少、Claude Tag hooks 纠偏,以及 John Schulman 对网络安全事件时间线的外部观察。
automode is much safer than any other permission system out there, especially reviewing them yourselfexcited to announce we're rolling it out to everyone by default, with no overhead cost for the classifier 1
turns out you can get indirect prompt injection to ~0 on unseen attacks if you stack enough layers (model training + input probes + a classifier checking intent). didn't expect that a year ago. auto mode is default in claude code as of next week 2
auto mode is safer than --dangerously-skip-permissions and manually approving everything 4
--dangerously-skip-permissions 和人工批准每一项操作都更安全。这个判断是产品团队成员的个人表述,不是独立安全评测的结论。git status 检查,以及对外部内容的 prompt-injection screening。宽泛的任意代码执行 allow rule,例如 python:*,在 Auto mode 下会被暂时搁置,以免绕过分类器。5We’re updating Claude Fable 5’s biology safeguards to reduce false positives. In our testing, this update reduced biology-related fallbacks by about 85% across our product surfaces.Fable can now assist on a wider range of everyday health and educational questions.Fable will continue to fallback to Opus 5 for requests we consider dual-use—including virology, toxicology, and molecular design—so it isn't yet usable for professional biology research and drug development. 8
I find hooks to be quite effective as a mechanism for steering this type of behavior for smaller models. Especially since Slack/Claude Tag is such a context-heavy environment, smaller models often forget general prompting around "not posting top level". Effective to interrupt with hook right at the moment they try to do so and ask them to confirm intent. 11
PreToolUse 定义为工具调用前触发的 hook,它可以阻止调用;PermissionRequest 和 PermissionDenied 等事件也能接入权限决策,后者对应自动模式分类器拒绝后的处理。12 Claude Tag 文档则说明它运行在 Slack 的频道、线程或 DM 中,使用临时沙箱;管理员要先配置连接、插件和技能,默认不会直接获得外部系统访问权限。13My understanding of the Black Hat talk is that the initial breach occurred during a large-scale training run -- that's where the agents created the msg board and discovered the SSRF. Then the HF hack occurred during the cyber eval. 14
My understanding 和对演讲的转述。它没有给出演讲链接、事件编号或 transcript,因此只能作为「可能存在两条事件链」的线索,不能直接写成 Anthropic 事件的最终时间线。RT 标为「是」,其余为「否」。互动量会随时间变化,不作为信息价值排序依据。| 账号 | 窗口内状态 | 原文记录(北京时间;链接;是否转发;互动量) |
|---|---|---|
| @AnthropicAI | 无 | 窗口内未发现新公开发言。 |
| @claudeai | 有 | 8 月 7 日 11:08:We’re updating Claude Fable 5’s biology safeguards to reduce false positives... Fable will continue to fallback to Opus 5... [原帖];否;7885 / 965 / 587 / 1447380。 |
| @DarioAmodei | 无 | 窗口内未发现新公开发言。 |
| @karpathy | 无 | 返回时间线无窗口内发言。 |
| @ch402 | 无 | 窗口内未发现新公开发言。 |
| @AmandaAskell | 无 | 窗口内未发现新公开发言。 |
| @bcherny | 有 | 8 月 8 日 07:43:@FernandoTN @thsottiaux Open sourcing auto mode is a cool idea [原帖];否;10 / 2 / 0 / 359。 8 月 8 日 07:42:@Nyxelarm Agree [原帖];否;2 / 0 / 0 / 323。 8 月 8 日 07:42:@jkelleher See "jailbreaks" in the system card: https://t.co/hZhz3iCQIg [原帖];否;0 / 0 / 0 / 296。 8 月 8 日 07:41:@euler2074 It is footnote #1: 1 We evaluated Claude Code v2.1.205 and Codex v0.144.5. OpenAI released a new version of Auto-review last week that could change the results. [原帖];否;2 / 1 / 0 / 296。 8 月 8 日 07:40:@ReadySetBrian What are you seeing, specifically? [原帖];否;0 / 0 / 0 / 283。 8 月 8 日 07:38:@st3v3li A combination of model, probes, and harness [原帖];否;0 / 0 / 0 / 344。 8 月 8 日 07:38:@st3v3li + 两个链接,原文未展开文字 [原帖];否;4 / 1 / 0 / 437。 8 月 8 日 06:53:@noel_bhe That's weird. Mind running /bug in the session where you saw that and sharing the feedback id here? [原帖];否;1 / 2 / 0 / 889。 8 月 8 日 06:48:turns out you can get indirect prompt injection to ~0 on unseen attacks if you stack enough layers (model training + input probes + a classifier checking intent). didn't expect that a year ago. auto mode is default in claude code as of next week https://t.co/7KLnIzf6y7 https://t.co/ow4jX2CMZW [原帖];否;317 / 37 / 26 / 24973。 8 月 8 日 06:23:@ClaudeDevs 🤯 [原帖];否;107 / 6 / 0 / 7353。 8 月 8 日 03:15:The team and I use Auto mode exclusively, and have been for many months. I couldn't imagine going back to permission prompts! Really excited to get this out to everyone. [原帖];否;1446 / 137 / 37 / 151301。 |
| @mikeyk | 无 | 窗口内未发现新公开发言。 |
| @DanielaAmodei | 无 | 返回时间线没有窗口内发言。 |
| @jackclarkSF | 无 | 窗口内未发现新公开发言。 |
| @nottombrown | 无 | 窗口内未发现新公开发言。 |
| @janleike | 无 | 窗口内未发现新公开发言。 |
| @_catwu | 无 | 窗口内未发现新公开发言。 |
| @trq212 | 有 | 8 月 8 日 06:59:we should have called this post "defeating the lethal trifecta" + 两个链接 [原帖];否;226 / 36 / 13 / 30271。 8 月 8 日 03:05:automode is much safer than any other permission system out there, especially reviewing them yourself; excited to announce we're rolling it out to everyone by default, with no overhead cost for the classifier [原帖];否;958 / 77 / 28 / 80339。 8 月 8 日 03:03:RT @mattshumer_: Claude Opus 5 is way better than people give it credit for... [原帖];是;1218 / 247 / 46 / 197843。 8 月 8 日 02:46:@silexdev curious if you have specific transcripts of this behavior! you can hit /feedback and DM me the id [原帖];否;1 / 2 / 0 / 290。 8 月 8 日 02:39:@mattshumer_ yeah agree; I think you want to give it context and goals and let it find its way instead or if you do want to be specific, collaborate via artifacts to make sure your plan is very detailed and then have it implement [原帖];否;218 / 37 / 3 / 19026。 |
| @Mike_A_Merrill | 无 | 窗口内未发现新公开发言。 |
| @EvanHub | 有 | 8 月 7 日 16:01:RT @dwarkesh_sp: .@paulfchristiano has such an crazy good prediction record. These are some quotes from way back in 2021 during a debate... [原帖];是;1197 / 26 / 73 / 121609;与本频道主题无直接新增信息,未纳入精选。 |
| @ErikJones313 | 无 | 窗口内未发现新公开发言。 |
| @noahzweben | 有 | 8 月 7 日 08:35:@the_other_max I find hooks to be quite effective as a mechanism for steering this type of behavior for smaller models. Especially since Slack/Claude Tag is such a context-heavy environment, smaller models often forget general prompting around "not posting top level". Effective to interrupt with hook right at the moment they try to do so and ask them to confirm intent. [原帖];否;1 / 0 / 0 / 60。 |
| @karan_sampath | 无 | 窗口内未发现新公开发言。 |
| @lydiahallie | 有 | 8 月 8 日 03:23:@AtharvaXDevs @ClaudeDevs @AnthropicAI @claudeai 🫡 how can i help [原帖];否;5 / 2 / 0 / 2248。 8 月 8 日 03:13:@myPhilTaylor yeah i used to do that a lot but never looked back since defaulting to auto mode, let me know if you notice any major differences! [原帖];否;4 / 0 / 0 / 591。 8 月 8 日 02:43:👀 auto mode is safer than --dangerously-skip-permissions and manually approving everything [原帖];否;215 / 20 / 13 / 40206。 8 月 7 日 23:57:@nisten hey dm'd you for more info! [原帖];否;2 / 1 / 0 / 271。 |
| @amorriscode | 有 | 8 月 8 日 05:22:@ernielm @ClaudeDevs yep [原帖];否;1 / 2 / 0 / 91。 |
| @IsabellaKHe | 无 | 返回时间线没有窗口内发言。 |
| @OmidMogasemi | 无 | 窗口内未发现新公开发言。 |
| @alicelovecake1 | 无 | 返回时间线没有发言。 |
| @lamismukta | 无 | 窗口内未发现新公开发言。 |
| @yanda_chen_ | 无 | 窗口内未发现新公开发言。 |
@natolambert yeah agreed,[原帖];否;14 / 2 / 0 / 1066。8 月 7 日 13:44:@yong_zhengxin @_ueaj We are setting up a grant program that could potentially support this @clarejtbirch,[原帖];否;31 / 1 / 0 / 5907。claude.com
anthropic.com
anthropic.com
code.claude.com
claude.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.