Accountability

Corrections

We log substantive corrections here. Minor typographical fixes are made silently.

DateArticleWhat changed
Thu Oct 22 2026 00:00:00 GMT+0000 (Coordinated Universal Time)PHP 8.5.12 fixes two critical pre-auth RCEs: update nowThe CISA KEV tracker. We checked the published catalogue file directly: catalogVersion 2026.10.21 at 14:04 UTC, neither CVE listed. A separate CISA alert telling admins to patch is a different document.
Thu Oct 22 2026 00:00:00 GMT+0000 (Coordinated Universal Time)PHP 8.5.12 fixes two critical pre-auth RCEs: update nowWiz's report contradicts our 14:12 entry on which bug is which. It labels CVE-2026-109618 an upload-progress unserializer use-after-free, the same description we reported as wrong. Both descriptions carry the same CVE. We have kept our entry and marked it disputed rather than silently reversing it.
2026-10-04Cisco SD-WAN Manager auth bypass CVE-2026-76504 is liveAn earlier version attributed an alert to the UK's National Cyber Security Centre. The linked alert is from New Zealand's NCSC, and it reports active exploitation rather than assessing further exploitation as highly likely; that assessment is NHS England's. We also removed a chart that plotted software version numbers as bar heights; the table of fixed releases carries that information.
Sun Oct 04 2026 00:00:00 GMT+0000 (Coordinated Universal Time)DevDay shipped the agent. The training pause stayedRewritten after the keynote. The 1 October update had left the piece as a preview of an event that had already happened. Rumours that were not announced (Aeon, GPT-6 Cyber, an assistant called o) are now labelled as such. Prices are taken from OpenAI's pricing page.
2026-10-04AI agent sandbox escape at OpenAI hits public wikiWe merged our separate report on the same incident into this article, added links to the researchers' report, their data dump, OpenAI's technical report and METR's investigation, and attributed the bypass details to the researchers rather than to a news summary. An earlier version said the Hugging Face agents tunnelled out through Artifactory; OpenAI's report describes Artifactory as the channel they used as a message board.

To report an error, email [email protected]. How we handle corrections is described in our editorial policy.