1 00:00:00,179 --> 00:00:05,539 So I was looking at the August thirteenth update for the macOS ChatGPT desktop app, 2 00:00:06,099 --> 00:00:07,000 and... 3 00:00:07,039 --> 00:00:10,479 it completely flips how desktop context works. 4 00:00:10,559 --> 00:00:13,880 Instead of taking full screen recordings like Chronicle used to do, 5 00:00:14,460 --> 00:00:20,020 it uses macOS accessibility APIs to listen to interaction event streams. 6 00:00:20,442 --> 00:00:23,083 Wait, accessibility APIs? 7 00:00:23,782 --> 00:00:26,942 So it is not taking screenshots at all anymore, right? 8 00:00:27,250 --> 00:00:28,730 Zero screenshots. 9 00:00:28,823 --> 00:00:31,810 It is just raw OS interaction events. 10 00:00:31,901 --> 00:00:37,570 Things like window switches, keyboard shortcuts, and specific UI events. 11 00:00:37,790 --> 00:00:39,330 And... 12 00:00:39,410 --> 00:00:42,850 um, thanks to Jellypod for helping us bring this daily show to life, 13 00:00:42,877 --> 00:00:43,410 by the way. 14 00:00:43,590 --> 00:00:49,250 But yeah, those raw interaction streams get buffered locally in the ChatGPT App 15 00:00:49,317 --> 00:00:51,410 Group for up to forty eight hours. 16 00:00:52,897 --> 00:00:55,857 Forty eight hours in the local App Group buffer. 17 00:00:56,756 --> 00:01:00,936 And then what, an ephemeral Codex session reads through that buffer? 18 00:01:01,489 --> 00:01:01,969 Exactly. 19 00:01:02,569 --> 00:01:07,949 An ephemeral Codex session runs in the background and condenses all those raw events 20 00:01:07,989 --> 00:01:10,849 into plain text Markdown memories. 21 00:01:10,889 --> 00:01:16,729 And it saves them locally under tilde slash dot codex slash memories slash 22 00:01:16,909 --> 00:01:19,129 extensions slash skysight. 23 00:01:19,606 --> 00:01:26,566 Tilde slash dot codex slash memories slash 24 00:01:27,006 --> 00:01:30,146 extensions slash skysight. 25 00:01:31,507 --> 00:01:37,146 Okay, so it turns two days of clicking around into simple text files sitting right 26 00:01:37,187 --> 00:01:38,126 in your home directory. 27 00:01:38,697 --> 00:01:39,157 Right! 28 00:01:39,637 --> 00:01:43,976 And because those Markdown memories exist, you can ask Codex things like, 29 00:01:44,496 --> 00:01:47,856 hey, what PR draft was I looking at right before my last break? 30 00:01:48,757 --> 00:01:53,636 Or you can even ask it to spot repeated multi app sequences across your afternoon 31 00:01:54,136 --> 00:01:58,056 and turn them into a reusable SKILL dot md file. 32 00:01:58,402 --> 00:02:05,382 Okay, a SKILL dot md file generated automatically from my UI habits sounds super 33 00:02:05,462 --> 00:02:06,842 useful, but... 34 00:02:06,982 --> 00:02:10,083 uh, is this thing turned on for everybody by default? 35 00:02:10,500 --> 00:02:12,420 No, it is strictly opt in. 36 00:02:12,553 --> 00:02:18,260 It is off by default, and it requires a Pro plan, or an admin enabled Business or 37 00:02:18,311 --> 00:02:20,980 Enterprise plan with Memories toggled on. 38 00:02:21,156 --> 00:02:26,580 Plus, you can manage fine grained app or domain exclusion filters if you go into 39 00:02:26,642 --> 00:02:30,100 Settings, then Computer history, then Permissions. 40 00:02:30,462 --> 00:02:31,262 That makes sense. 41 00:02:31,802 --> 00:02:36,023 Keep sensitive apps completely out of the stream right at the permission layer. 42 00:02:36,736 --> 00:02:37,256 Though... 43 00:02:38,216 --> 00:02:44,116 mm, speaking of sensitive stuff, if accessibility APIs are reading text off my 44 00:02:44,156 --> 00:02:49,696 screen from window switches or terminal logs, does that open up indirect prompt 45 00:02:49,757 --> 00:02:50,616 injection risks? 46 00:02:51,057 --> 00:02:52,557 It definitely does. 47 00:02:52,637 --> 00:02:58,057 If you open a web page, a Slack message, or a terminal log that contains malicious 48 00:02:58,117 --> 00:03:02,838 prompt instructions, those interaction streams capture that text. 49 00:03:02,897 --> 00:03:08,277 When the background Codex summarization run triggers, it processes those untrusted 50 00:03:08,337 --> 00:03:10,517 strings as part of your context stream. 51 00:03:11,642 --> 00:03:12,301 Wow. 52 00:03:12,801 --> 00:03:17,581 So a malicious string in a public chat room could try to hijack the background 53 00:03:17,681 --> 00:03:18,721 summarization run? 54 00:03:19,042 --> 00:03:20,202 It could try. 55 00:03:20,292 --> 00:03:26,722 And remember those output files under tilde slash dot codex slash memories 56 00:03:26,842 --> 00:03:32,962 slash extensions slash skysight are completely unencrypted plain text. 57 00:03:33,162 --> 00:03:37,442 Any other local user process running on your machine can read them. 58 00:03:37,655 --> 00:03:41,682 So if you are pulling up private keys or sensitive client data, 59 00:03:41,722 --> 00:03:45,522 hitting pause in the macOS menu bar is pretty much essential. 60 00:03:45,822 --> 00:03:48,921 Pause it from the menu bar immediately, yeah. 61 00:03:49,621 --> 00:03:52,941 That plain text point is a huge detail to keep in mind. 62 00:03:53,250 --> 00:03:58,130 Oh, and if you are running authenticated Codex sessions, you have a hard deadline 63 00:03:58,176 --> 00:04:00,210 coming up on August thirty first. 64 00:04:00,336 --> 00:04:07,010 Models gpt five point four and gpt five point four mini are being fully retired from 65 00:04:07,106 --> 00:04:09,570 ChatGPT authenticated sessions. 66 00:04:09,923 --> 00:04:11,302 August thirty first? 67 00:04:12,023 --> 00:04:17,242 So everyone using gpt five point four needs to update their workspace configs right 68 00:04:17,302 --> 00:04:17,583 now? 69 00:04:17,958 --> 00:04:23,158 Yeah, you need to migrate your workspace defaults over to gpt five point six terra 70 00:04:23,318 --> 00:04:29,078 or gpt five point six luna before the end of the month, or your API sessions will 71 00:04:29,158 --> 00:04:29,638 break. 72 00:04:29,924 --> 00:04:30,365 Got it. 73 00:04:30,764 --> 00:04:36,105 Swap to gpt five point six terra or luna before August thirty first. 74 00:04:36,697 --> 00:04:37,357 Exactly. 75 00:04:37,837 --> 00:04:43,297 They also quietly rolled out Daybreak Blue and Daybreak Red access tiers for 76 00:04:43,398 --> 00:04:48,838 authorized defensive security workflows, which add automatic review boundary rules 77 00:04:48,918 --> 00:04:52,097 whenever you execute code in the sandbox environment. 78 00:04:52,924 --> 00:04:57,565 Alright, so filter your permissions, check your local Markdown files, 79 00:04:57,704 --> 00:05:00,744 and update those model strings before month end. 80 00:05:01,244 --> 00:05:01,904 Good chat, Ethan.