1 00:00:00,000 --> 00:00:00,880 Morning, everyone. 2 00:00:00,983 --> 00:00:04,640 Thanks to Jellypod for helping make this daily show a reality. 3 00:00:04,773 --> 00:00:10,240 Today is Codex Security 0.2.0, published on October 6. 4 00:00:10,453 --> 00:00:16,560 If you caught our earlier security episode on 0.1.29, with the markdown severity 5 00:00:16,620 --> 00:00:21,200 rubrics and custom patch validation prompts, this release builds on that. 6 00:00:21,380 --> 00:00:25,040 The threat model behind those policies now gets saved too. 7 00:00:25,342 --> 00:00:26,542 Here's the exact change. 8 00:00:27,082 --> 00:00:31,162 Threat models are now saved with Standard, Deep, and Diff scan results, 9 00:00:31,642 --> 00:00:33,362 and with generated security policies. 10 00:00:33,802 --> 00:00:38,422 That's pull request 1133, and the release marks it as a breaking change. 11 00:00:39,002 --> 00:00:40,782 The headline command is export. 12 00:00:41,222 --> 00:00:45,642 You run codex security export, with the artifact flag set to threat model. 13 00:00:46,162 --> 00:00:50,422 Or, if you're in the TypeScript SDK, there's an exportArtifact helper. 14 00:00:51,062 --> 00:00:55,122 The release notes say you can export a saved model without another model call. 15 00:00:55,375 --> 00:00:56,815 That last part is the point. 16 00:00:56,955 --> 00:01:01,455 The notes state it plainly: no additional model invocation. 17 00:01:01,599 --> 00:01:06,415 What follows is my interpretation, not something the release notes spell out. 18 00:01:06,595 --> 00:01:12,015 The threat model becomes something you can store alongside a scan and pull into a PR 19 00:01:12,075 --> 00:01:17,295 summary, an audit doc, or a review checklist without paying for another model call 20 00:01:17,348 --> 00:01:18,255 to regenerate it. 21 00:01:18,435 --> 00:01:23,855 And because it's saved with that scan's results, the text you export is the text 22 00:01:23,887 --> 00:01:25,935 that existed when the scan ran. 23 00:01:26,358 --> 00:01:27,638 Let me walk through the workflow. 24 00:01:28,158 --> 00:01:30,418 First, run a scan in one of the three modes. 25 00:01:30,718 --> 00:01:36,418 Something like npx, then the at openai slash codex security package, 26 00:01:36,958 --> 00:01:40,498 then scan, with the target flag pointing at your source directory. 27 00:01:40,978 --> 00:01:46,378 Threat modeling is part of Standard, Deep, and Diff, so there's no extra switch to 28 00:01:46,418 --> 00:01:46,758 turn on. 29 00:01:47,438 --> 00:01:49,038 Once that finishes, you export. 30 00:01:49,578 --> 00:01:55,558 Same package through npx, the export command, the artifact flag set to threat model, 31 00:01:56,158 --> 00:01:57,978 and the scan flag with your scan ID. 32 00:01:58,827 --> 00:02:04,167 In the SDK, you create a client, then await client dot exportArtifact, 33 00:02:04,387 --> 00:02:07,667 passing the scan ID and the artifact name, threat model. 34 00:02:08,267 --> 00:02:09,487 That returns the saved model. 35 00:02:10,287 --> 00:02:14,927 I'd check the SDK types for the exact signature before wiring it into a pipeline, 36 00:02:15,447 --> 00:02:18,327 because the shape of what comes back matters a lot in this release, 37 00:02:18,787 --> 00:02:19,747 and that's the next part. 38 00:02:20,042 --> 00:02:24,442 Right, because there are three upgrade notes that affect anything reading these 39 00:02:24,509 --> 00:02:25,002 files. 40 00:02:25,149 --> 00:02:31,002 First, policy generation now writes a file called threatmodel.md, 41 00:02:31,122 --> 00:02:36,602 all lowercase, instead of THREAT_MODEL.md in capitals. 42 00:02:36,762 --> 00:02:41,482 The notes tell you to use the returned path rather than hard coding either name. 43 00:02:41,642 --> 00:02:48,602 Second, on the SDK side, check SecurityPolicyDraft.threatModelPath for null 44 00:02:48,671 --> 00:02:50,042 before you open the file. 45 00:02:50,242 --> 00:02:53,882 The notes don't say when it's null, so just guard against it. 46 00:02:54,087 --> 00:02:57,674 Third, saved models can come back in two forms. 47 00:02:57,854 --> 00:03:03,714 One is Markdown, where you get a format field set to markdown plus a content field. 48 00:03:03,804 --> 00:03:06,554 The other is the previous structured format. 49 00:03:06,661 --> 00:03:08,554 So a reader has to handle both. 50 00:03:08,734 --> 00:03:13,594 The good news is that existing models stay readable and exportable without 51 00:03:13,666 --> 00:03:14,714 migration. 52 00:03:14,814 --> 00:03:18,234 You don't have to rewrite old scans to use the new export. 53 00:03:18,630 --> 00:03:22,070 Now the caveats, because a few of them could bite you in CI. 54 00:03:22,650 --> 00:03:28,470 The one I'd read twice: Codex Security no longer masks diagnostic text that looks 55 00:03:28,510 --> 00:03:29,310 like credentials. 56 00:03:29,750 --> 00:03:31,710 That's pull request 1179. 57 00:03:32,250 --> 00:03:37,590 The notes say CLI output, plus saved error, publication, and patch risk summaries, 58 00:03:38,050 --> 00:03:41,930 can contain sensitive values, and that you should review them before sharing. 59 00:03:42,270 --> 00:03:45,530 So if you were leaning on the tool to scrub those logs, you need to do your own 60 00:03:45,610 --> 00:03:48,330 scrubbing now, before anything gets posted or archived. 61 00:03:48,708 --> 00:03:54,628 Second, if you run the findings service, clients must send a Content Type header of 62 00:03:54,701 --> 00:04:01,108 application slash json to POST slash v1 slash bulk slash 63 00:04:01,197 --> 00:04:07,428 findings and POST slash v1 slash dedupe dash groups. 64 00:04:07,575 --> 00:04:11,588 Sorry, I mean the dedupe groups endpoint, under v1. 65 00:04:11,788 --> 00:04:18,228 A missing or different content type returns HTTP 400, invalid_request. 66 00:04:18,588 --> 00:04:25,348 A charset parameter is accepted, so application slash json with a charset is fine. 67 00:04:25,568 --> 00:04:29,028 Any script that posts a bare body will break on upgrade. 68 00:04:29,322 --> 00:04:29,742 Two more. 69 00:04:30,422 --> 00:04:35,802 SDK types and CLI schemas now accept any nonempty string for reasoning effort, 70 00:04:36,162 --> 00:04:36,942 not a fixed list. 71 00:04:37,342 --> 00:04:40,962 So if your integration validates against a hard coded set of values, 72 00:04:41,402 --> 00:04:41,822 update it. 73 00:04:42,302 --> 00:04:45,302 The value still has to be supported by Codex and your model provider, 74 00:04:45,362 --> 00:04:45,522 though. 75 00:04:46,422 --> 00:04:51,462 And if you export results as CSV, older exports may have ambiguous leading 76 00:04:51,522 --> 00:04:54,362 apostrophes that can't be recovered from the CSV alone. 77 00:04:54,942 --> 00:04:59,322 If you need those exact values, export the original saved results as JSON. 78 00:04:59,583 --> 00:05:01,903 Now the smaller quality of life changes. 79 00:05:02,116 --> 00:05:07,583 Scans now check for exposed credentials in source code, including unused code and 80 00:05:07,663 --> 00:05:12,063 tests, and they distinguish suspected exposures from placeholders. 81 00:05:12,207 --> 00:05:17,503 That check runs offline, and it doesn't try any discovered credential against a 82 00:05:17,553 --> 00:05:18,063 service. 83 00:05:18,207 --> 00:05:21,743 That's worth knowing if your security team has rules about that. 84 00:05:22,050 --> 00:05:25,670 On the command side, the model and effort flags are now shared. 85 00:05:25,990 --> 00:05:29,370 You can use them when generating patches, validating findings, 86 00:05:29,710 --> 00:05:33,090 verifying fixes, or matching and comparing saved scans. 87 00:05:33,690 --> 00:05:38,350 Deep Scan workers and patches generated during a scan also respect your selected 88 00:05:38,410 --> 00:05:38,950 service tier. 89 00:05:39,410 --> 00:05:44,730 Patch, validation, and fix verification commands now respect custom model providers 90 00:05:44,830 --> 00:05:46,250 and their authentication settings. 91 00:05:46,950 --> 00:05:51,230 Amazon Bedrock users get clearer authentication errors and cost estimates for 92 00:05:51,270 --> 00:05:52,310 Daybreak Blue and Red. 93 00:05:52,910 --> 00:05:58,490 And eligible OpenAI API key users can select a Cyber access program for each scan. 94 00:05:58,625 --> 00:06:00,065 Coverage expanded too. 95 00:06:00,305 --> 00:06:07,185 Scans now pick up C++ headers ending in dot h h and dot h x x, 96 00:06:07,311 --> 00:06:13,985 server rendered templates in EJS, ERB, and PHTML, and Vyper 97 00:06:14,042 --> 00:06:14,865 source files. 98 00:06:15,038 --> 00:06:19,505 Files that used to be silently skipped may now produce findings, 99 00:06:19,558 --> 00:06:22,385 so expect your first scan after upgrading to look different. 100 00:06:22,677 --> 00:06:24,537 And two behavior changes I like. 101 00:06:24,977 --> 00:06:30,077 On Unix, scans check that the sandbox works before starting paid model calls, 102 00:06:30,437 --> 00:06:33,757 so a broken sandbox gets caught early instead of mid run. 103 00:06:34,337 --> 00:06:38,217 And if you cancel follow up work, completed results are kept. 104 00:06:38,458 --> 00:06:41,418 One more item, outside Codex Security. 105 00:06:41,518 --> 00:06:47,218 Codex CLI 0.160.1 shipped on October 5. 106 00:06:47,338 --> 00:06:53,818 It preserves SYSTEMROOT, TEMP, and TMP when launching remote stdio MCP 107 00:06:53,898 --> 00:06:57,978 servers that have explicitly configured remote environment variables. 108 00:06:58,122 --> 00:07:02,298 That lets Unix hosts keep the Windows executor's startup environment. 109 00:07:02,538 --> 00:07:08,538 If you run a Unix host driving a Windows executor, install it with npm install dash 110 00:07:08,778 --> 00:07:14,218 g at openai slash codex at 0.160.1. 111 00:07:14,450 --> 00:07:16,410 So the practical checklist is short. 112 00:07:16,870 --> 00:07:21,490 Upgrade, scan, export the threat model once to see which format you get, 113 00:07:21,970 --> 00:07:26,230 then fix any reader that assumes the old file name or skips the null check. 114 00:07:26,670 --> 00:07:28,610 Then audit where your CI logs go. 115 00:07:29,030 --> 00:07:29,670 That's it for today.