2026-10-12
Top stories
- Zod 4.6.6 published to npm with a malicious preinstall script A malicious
preinstallscript shipped in zod 4.6.6 and @zod/mini 4.6.6, running on plain npm install without import. - Cyberattack on Danfoss refrigeration software heats fridges in hundreds of French supermarkets Attackers took over Danfoss refrigeration management software and ordered fridges in hundreds of French supermarkets to heat.
- Torvalds publishes a page-cache patch series written with Claude Code and says so in the patch message Torvalds posted an unmerged page-cache patch series and disclosed in the patch message that he wrote it with Claude Code.
- HashiCorp archives Boundary Community Edition with no further security patches planned HashiCorp archived the Boundary Community Edition repository and plans no further bug fixes or security patches.
- Sakana AI reports recursive self-improvement from multi-agent self-supervision, with uneven transfer Sakana AI reports 1.2x to 1.6x gains per output token from multi-agent self-supervision, but below 1x on SWE-bench Verified.
Security
- Flask session key recovered from its own cookie yields RCE across Google DevSite domains A Flask secret key stored inside its own session cookie gave pickle RCE across Google DevSite domains, paid at 90,000 dollars.
- TinaCMS CLI code injection via unescaped Git branch name in generated client An unescaped Git branch name interpolated into generated TinaCMS client code executes during preview builds.
- GitLab patches CVSS 10 unauthenticated RCE in the CI/CD regular expression compiler Self-managed GitLab 19.2 to 19.4 allows unauthenticated code execution through a crafted CI/CD regex until patched to 19.4.2, 19.3.5 or 19.2.8.
- php-fpm IPv6 allow list matched only 96 of 128 bits for eleven years (CVE-2026-91768) A memcmp that compared 12 of 16 bytes let php-fpm admit any IPv6 client in the same /96 as an allowed address, including all of ::/96 for ::1.
Outages
AI
- Anthropic finds small open models near the capability to self-replicate across IoT devices Anthropic red-teamers show small open models setting up inference on unfamiliar devices, with GLM-5.2 completing replication in a test fleet.
- Meta releases Muse Wonder and Muse Wonder Full for agentic and coding work Meta's new Muse Wonder model targets long-running agents and coding, and is available now through the Meta Model API and Muse Code.
Agentic coding
- QGIS merges a policy prohibiting AI contributions QGIS merged an enhancement proposal prohibiting AI contributions, citing maintainer review capacity rather than code quality.
- Codex CLI starts a persistent self-updating background daemon by default Codex CLI 0.157.0 starts a 374 MB self-updating background daemon that survives closing every Codex window.
Languages and runtimes
Infrastructure
Engineering posts
Hacker News
- Yandex Cloud outage and a collapse in Yandex bot traffic drive the day's largest threads Three HN threads on a reported Yandex Cloud outage drew about 1,350 points, with no primary source readable here.
- HN debates whether human review, not model capability, is now the binding constraint A 940-comment HN thread on a Karpathy post argues human code review, not model capability, now limits shipping.
- Passkey criticism draws technical dissent over whether password managers restore user control HN splits on passkeys, with commenters arguing KeePassXC and Bitwarden keep user control over the credentials.