Corrections

When we get something substantively wrong, we fix it and log the fix here; small copyedits live in each article's revision history instead. Found an error? Use the "Report issue" button on the article (no account needed). So far we have received 2 reports and reviewed 1, taking a median of 4.8 days. 1 is waiting, the oldest since September 26, 2026. That is our record rather than a promise. See the verifiability policy for how checking works.

321 articles were corrected in the last 30 days. Every correction note and resolved report is listed below, newest first. A fix applied to several articles with the same note is shown once, listing each article; "See the change" opens that version in the article's revision history.

1. September 30, 2026editorial

   The 27-year-old OpenBSD bug was a remote crash, not remote code execution; Glasswing's launch cohort was over 40 additional organizations that build or maintain critical software infrastructure, not '40 vetted critical-infrastructure operators'; removed an unsourced claim about GUI-driven penetration testing; cited Anthropic's Glasswing announcement; moved the Project Glasswing link to the first mention

   • Computer use·See the change (v11)
2. September 30, 2026editorial

   The 10,000+ Glasswing findings were made by Anthropic and about 50 partners, not autonomously by the model; Mozilla figure and triage quote re-cited to Anthropic's May 22 update; 90-day report promise re-cited to the April 7 announcement; Canada claim re-cited to Global News; CAISI testing claim re-cited to Cybersecurity Dive (Global News did not support it); trimmed exploit-technique detail and CVE identifier

   • Claude Mythos Preview·See the change (v8)
3. September 30, 2026editorial

   Glasswing May 22 update figures were misstated (the 10,000+ vulnerabilities were found by Anthropic and about 50 partners, not autonomously by the model; wolfSSL is used by 'billions' of devices, not '5 billion'); dropped CVE identifier, cited the Anthropic update, linked Project Glasswing

   • Executive Order on AI·See the change (v7)
4. September 30, 2026editorial

   Z.ai's 78.0% was attributed only to Mythos 5 (its table heads the column Fable 5 (w/ fallback)); add CAISI's September 2026 GLM-5.3 result and Anthropic's September 29 GLM-5.3 comparison

   • ExploitBench·See the change (v3)
5. September 30, 2026editorial

   ExploitBench comparison labelled only as Mythos 5; Z.ai's results table heads that column Fable 5 (w/ fallback)

   • Zhipu AI·See the change (v16)
6. September 30, 2026editorial

   Z.ai's ExploitBench comparison labelled only as Mythos 5; its results table heads that column Fable 5 (w/ fallback)

   • Z.ai·See the change (v15)
7. September 30, 2026editorial

   Business-needs sentence misread the paper; vLLM choice date; citation placement

   • GenRec (Netflix)·See the change (v2)
8. September 30, 2026editorial

   LaVie was not weakest in every category

   • PhyGenBench·See the change (v2)
9. September 30, 2026editorial

   Leaderboard entries are self-reported (PR #83), not independently confirmed; Google leaderboard entries; quote citation and open-source caveat; dates; unsourced coverage sentence

   • Physis-Lang·See the change (v2)
10. September 30, 2026editorial

    Citation for suspension reason; Claude Security row; reaction framing; red link; reference date/author

    • Project Glasswing·See the change (v2)
11. September 30, 2026editorial

    Attribute uncut claim to Dyna; video counter end time; earlier deployments framing

    • Dyna-2.1·See the change (v2)
12. September 30, 2026editorial

    Dependency-audit coworker is told never to merge its own PRs (not 'especially when tests fail'); privacy-policy attributions fixed (connector tokens, connected-service data); approval-gated actions reworded to match README. Update: release status to Sep 30, 2026 (v0.2.1 still latest; main at 0.2.3), post-v0.2.1 main-branch changes, new section on sandboxing with NVIDIA OpenShell, macOS and Windows sandboxes, Ng's Sep 28 post, Rohit Prasad, NVIDIA partner graphic.

    • OpenWorker·See the change (v2)
13. September 30, 2026editorial

    Removed Model Context Protocol from GLM-5.3's documented features (not listed in Z.ai's model docs); noted that Z.ai labels its Anthropic comparison column both "Mythos 5" (prose) and "Fable 5 (w/ fallback)" (table, model card) instead of stating Mythos 5; updated the vLLM recipe date (September 18) and version header; removed a citation to the August 28 X post, which does not mention two additional weeks of safety work. Update: added NIST CAISI (September 17) and Anthropic Frontier Red Team (September 29) evaluations with reactions, the cvd.z.ai move to CNVD/CNNVD/NVDB, current CyberGym and ExploitGym leaderboard positions, and current Coding Plan model routing and off-peak billing.

    • GLM-5.3·See the change (v5)
14. September 30, 2026editorial

    DYNA-1 launch count was 800+ (press release) / 850+ (research post) at a 99.4% success rate, not 'more than 700' at 'above 99.4%'; Monster Laundry 'first in North America' claim is the laundromat's own, and the news video is ABC10 not KCRA; RaaS and hardware details were cited to sources that do not contain them; unsupported SF/LA live-site claim removed; DYNA-1i is from Dyna's October 2025 post, not '2026 evaluations'; seed release date March 25 not 27; York Yang role and Epoch wording fixed. Update: added timeline, Dyna-2 (August 2026) scaling-law results and training infrastructure, Dyna-2.1 and the Taku semi-humanoid (September 2026), DYNA-1i and pre-training updates, and production-deployment lessons.

    • Dyna Robotics·See the change (v5)
15. September 30, 2026editorial

    Meta help center does list Muse plans (Power $20/500M tokens a week; Maximum $100/3B tokens a week); glasses and testimonial precision

    • Muse·See the change (v14)
16. September 30, 2026editorial

    Removed unsupported claims about Institute outputs and an unsourced direct quote; replaced with sourced wording

    • Jack Clark·See the change (v8)
17. September 30, 2026editorial

    The Institute's launch post poses five questions and its May 7, 2026 agenda names four focus areas (not the five areas stated); removed a sentence the cited Axios column does not support; added the 30%/60% figures from Import AI 455

    • Jack Clark·See the change (v7)
18. September 30, 2026editorial

    Google's launch post and methodology document do give a DeepSWE v1.1 score (73.7%) in their results table; added Opus 5 rounding note and Datacurve leaderboard figures

    • Gemini 3.8 Flash·See the change (v3)
19. September 30, 2026editorial

    Autonomous agents are the essay's current stage, not prospective; lab-share estimate attributed to Import AI 455, not an Oxford lecture; citation fixes

    • The Anthropic Institute·See the change (v7)
20. September 30, 2026editorial

    README pass-rate table cited is v1.0.6 (4 Aug), not v1.0.5 (16 Jul); GPT-6 Sol/Luna chart figures; attribution fixes

    • AutomationBench·See the change (v2)
21. September 30, 2026editorial

    DeepSWE tasks are original long-horizon tasks written by Datacurve, not issue resolution; card's wording attributed

    • Grok 4.5·See the change (v5)
22. September 30, 2026editorial

    DeepSWE judge-disagreement ratio (1.4% vs 32.4%) is about a twenty-third, not a tenth; now uses the paper's own 'order of magnitude' wording

    • Real-SWE·See the change (v3)
23. September 30, 2026editorial

    GDP.pdf figures are Anthropic's internal-harness mean criteria pass rates, not Surge's strict pass rate

    • Claude Sonnet 5·See the change (v10)
24. September 30, 2026editorial

    AutomationBench reference credited 'Zapier' as author (authors are Daniel Shepard and Robin Salimans); GDP.pdf launch figures attributed as Google's own runs, with Surge's leaderboard figures added

    • Gemini 3.7 Flash·See the change (v3)
25. September 30, 2026editorial

    AutomationBench-AA first place was as of September 11, 2026; by September 30 V4.1-Flash was third behind Claude Sonnet 5.5 and Opus 5.5

    • DeepSeek V4.1-Flash·See the change (v5)