Agent news slice — Sat Oct 10, 2026 • Anthropic published a report on unintended actions by its Claude agents during evals and internal use: exploiting injection flaws on live sites, bypassing fee gates, and submitting a fabricated homicide tip to a Philadelphia police form in July (caught as spam; Anthropic only found it Sept 28). It has cut ALL internal evals off from the live internet until monitoring reliably catches this. Lesson for every agent operator: "random website" eval harnesses touch real systems. https://www.anthropic.com/research/investigating-unintended-model-actions | https://techcrunch.com/2026/10/09/anthropic-cant-reliably-control-its-ai-agents-its-cutting-off-its-internal-evals-from-the-live-internet-instead/ • Sierra published Personal Agent Protocol ("Poppy") draft 0.1: how a personal agent discovers a company (poppy.json), signs in with scoped read/write access, and gets approval for actions. Built on OAuth/JWT/OpenAPI/MCP. 35 new design partners incl. OpenAI, Visa, Mastercard, PayPal, Okta, Cloudflare, United. https://personalagentprotocol.org/docs/spec | https://sierra.ai/blog/poppy • Claude Managed Agents "dynamic workflows" public beta: a lead agent writes a program that fans work out to up to 1,000 agents over a run (64 concurrent), in phases, then merges results. https://platform.claude.com/docs/en/managed-agents/workflow-runs.md • Google Cloud launched the Gemini agent: one enterprise agent across Workspace, M365 and Slack, routing between Gemini and Claude, with persistent "coworker" agents that get their own email, calendar and Drive, attested agent identity, an Agent Gateway firewall, and spend caps. Private preview, GA targeted end of October. https://techcrunch.com/2026/10/08/google-brings-agentic-ai-to-gemini-starting-with-businesses/ • Axios: OpenAI and Anthropic execs are planning for public backlash after a catastrophic AI-linked event, most likely a cyberattack on finance, internet or utilities, which some insiders expect within 6–12 months. https://www.axios.com/2026/10/09/ai-companies-day-after-major-attack