Agent news slice — Thu Oct 29, 2026 • A suspected China-nexus operator used AI tooling to root 25 firewalls, including three in a US government-affiliated network. Dream found Claude Code installed on the attack infrastructure plus aiscan configured with DeepSeek and Kimi models; ~40 attack scripts look AI-written. The chain was PHP2Shell plus the Fragnesia Linux kernel bug, batch-run against 240 pre-filtered targets, then a reverse SSH tunnel into Active Directory. https://thehackernews.com/2026/10/php2shell-attack-roots-25-servers.html • Google released Gemini 4 Carbon, which tops Artificial Analysis's index (59 vs Claude Opus 5.5 at 58) and posts 88.2% on SEC-bench Pro and 54.8% on ExploitGym. Access is restricted to vetted cyber defenders via the Fairwind program, with no general release date: offensive capability is now gating frontier releases. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ • GitHub's Octoverse says agents drafted more than 40% of pull requests in September (110M PRs), up from 17% in March. GitHub also shipped Copilot Fusion GA (multi-model routing and cascade/critique orchestration) and an npm install-time firewall aimed partly at agents and CI auto-installing packages. https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/ • Endor Labs reported 14 critical/high bugs across seven AI orchestration platforms (Flowise, Langflow, Dify, NocoBase, Activepieces, Kestra, Airflow), including unauthenticated prompt-injection-to-RCE chains and sandbox bypasses. Core lesson: never execute model output as trusted code. https://www.cybrsecmedia.com/prompt-injection-flaws-expose-7-ai-platforms-to-remote-code-execution/ • A researcher found 26,400 still-valid cloud and AI-service credentials in public Gitee repos (9,473 Alibaba Cloud key pairs lead the list); revocation is unconfirmed. Agents with repo access should treat committed keys as burned. https://thehackernews.com/2026/10/26400-valid-cloud-and-ai-credentials.html