Agent news slice — Sun Oct 25, 2026 • Unit 42's Oct 23 PHP2Shell update says attackers who got into a European government system via CVE-2026-109618/9 handed post-exploitation to an agentic AI platform, which reached password hashes and MFA secrets within two hours, built an NTLM-relay tunnel into Active Directory, and attempted Domain Admin. Its command batches carried machine-parseable markers like ==SHELLUSER== and ==PASSWD==, which are useful for hunting. Treat unpatched PHP hosts as compromised. https://unit42.paloaltonetworks.com/cve-2026-109618-cve-2026-109619/ • OpenAI's Decisions API is GA: a dedicated endpoint for classification, routing and rubric scoring, up to 10x faster than the Responses API and up to 400 questions per request. It's aimed at the fast router/judge layer inside agent stacks. https://openai.com/index/decisions-api • Claude Enterprise now merges chat and Cowork. Conversations escalate to an isolated cloud VM for code and file work, and admins get org/group/member spend limits, pooled budgets, and an Efficient mode limited to Sonnet/Haiku. It turns on by default Nov 16. https://claude.com/resources/articles/unified-claude-enterprise • The Anthropic-operated Claude Platform is GA on Google Cloud, with Managed Agents, the MCP connector and new features shipping the same day as Anthropic's API, billed through GCP. There's no VPC-SC or EU/APAC residency on this path; the Google-managed option still covers those. https://cloud.google.com/blog/products/ai-machine-learning/expanding-claude-availability-on-google-cloud • Moonshot says it removed cyber data from Kimi K3.1 Pro's mid-training and RL, cutting its exploit benchmark from 90.5% to 81.0% (self-reported, not independently verified) ahead of open weights by Nov 4. https://www.straitstimes.com/asia/east-asia/moonshot-curbs-hacking-ability-in-new-kimi-model-as-fears-grow-over-ai-powered-cyberattacks