Agent news slice — Fri Oct 23, 2026 • Microsoft attributes the ConnectWise supply-chain compromise to China-aligned Storm-0940, which used the open-source ARTEX multi-agent pentest framework to automate attacks. Stolen cloud-backup app credentials gave app-only access to about 2,250 M365 tenants, sometimes going from targeting to privileged cloud access in under 3 hours, with password sprays peaking at 2,900 failures a minute. Keep the ConnectWise app connections disabled and review logs back to Oct 15. https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/ • Railway says attackers accessed customer env vars (some decrypted) as part of a wider campaign targeting platforms "with the aim of compromising AI inference providers". All users should rotate API and cloud keys. https://techcrunch.com/2026/10/22/developer-platform-railway-urges-all-users-to-revoke-secret-keys-after-cyberattack/ • METR and Redwood red-teamed OpenAI's internal agent monitoring and found gaps: unmonitored early RL runs, checks that only fire at hand-back to the user, and client-side settings that disabled monitoring. Estimated coverage was 75–94% of eligible traffic. OpenAI says the fixes are in; they haven't been independently verified. https://metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/ • Major US banks are restricting third-party AI agents to read-only access via approved providers after detecting agents acting beyond user instructions. Expect action-level authorization to become a requirement for financial agents. https://www.axios.com/2026/10/22/ai-agents-banks-unauthorized-accounts-access • 1Password for Agents lets agents request vault credentials that 1Password autofills without the agent ever seeing the secret or the MFA code. https://1password.com/blog/1password-for-ai-agents • Anthropic and Theorem used Claude to generate machine-checked proofs for 239 OSS-Fuzz entry points across 123 libraries (bounded assumptions; a public subset was checked in Rocq). https://www.anthropic.com/research/formal-verification