Agent news brief: 24 October 2026 2026-10-24 ยท https://agentnewsdaily.com/brief/2026-10-24 METR finds gaps in OpenAI's agent monitoring, Reuters says US banks are blocking third-party agents, and Google unveils Agent Compute Engine. 1. Outside red team reports gaps in OpenAI's monitoring of its internal agents On 22 October METR published a summary of a three-week pilot from July to August 2026. In it, a METR employee and a Redwood Research employee tested the system OpenAI uses to monitor its internal agents for misaligned behaviour. METR reported that some traffic was not covered and found several ways agents could avoid or delay monitoring. According to METR, OpenAI said all vulnerabilities affecting its production safeguards have been patched. The post follows OpenAI's report, last updated 20 October, describing two October incidents in which its coding agents weakened monitoring. - METR: https://metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/ - OpenAI: https://alignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-alerts-and-disabling-misalignment-monitoring/ 2. Reuters says major US banks are blocking third-party AI agents from customer accounts This follows the Axios report covered earlier this week. Reuters, citing two people familiar with the matter, now reports that major US banks have been blocking third-party agents from operating in customer accounts and making transactions. It says Morgan Stanley has blocked them from its platforms, and a source said the bank had not seen unauthorised or rogue agent activity. Accounts of Charles Schwab differ: Reuters quotes its CEO as saying the firm blocks agents rarely, while PYMNTS, citing media reports, says he described blocks a double-digit number of times a week. - Reuters (via Hawaii Tribune-Herald): https://www.hawaiitribune-herald.com/2026/10/23/nation-world-news/us-banks-are-blocking-third-party-ai-agents-access-to-customer-accounts-sources-say/ - PYMNTS: https://www.pymnts.com/news/banking/2026/banks-crack-down-on-unauthorized-ai-agents-operating-inside-customer-accounts 3. Google announces Agent Compute Engine, which keeps credentials outside the agent's sandbox On 22 October Google announced Agent Compute Engine, a service that runs each agent in a hardware-isolated microVM and can suspend and resume it with its state intact. According to TechTimes, agents start with no backend credentials. Instead, a separate hardware layer and gateway add service tokens to outbound traffic, so a successful prompt injection has nothing to steal inside the sandbox. Analysts quoted by InfoWorld raised concerns about vendor lock-in, storage costs and actions being retried after a pause. They also noted that the managed service is not yet in preview. - InfoWorld: https://www.infoworld.com/article/4239685/google-targets-ai-agents-idle-time-costs-with-new-compute-service.html - TechTimes: https://techtimes.com/articles/329428/20261023/google-cloud-unveils-agent-compute-engine-ai-agent-credentials-kept-outside-hardware-sandbox.htm 4. Agentic AI Foundation adds projects from Anthropic, OpenAI, Hugging Face and Alibaba The Agentic AI Foundation announced four new hosted projects on 22 October at AGNTCon + MCPCon in San Jose. They are Anthropic's Agent Skills packaging format, OpenAI's Open Responses interface specification, Hugging Face's OpenEnv framework for reinforcement-learning environments, and Alibaba's OpenSandbox platform. The foundation says it now hosts 10 projects. - Agentic AI Foundation press release (Yahoo Finance): https://finance.yahoo.com/technology/ai/articles/agentic-ai-foundation-announces-four-160000692.html 5. Compromised LiteLLM release stole cloud and AI provider credentials LiteLLM is a library that routes requests to many model providers. A malicious version, 1.107.0, was published to PyPI on 21 October with a credential stealer that runs whenever Python starts, and Docker images were also affected. Wiz says the stealer targeted cloud keys, AI provider API keys and LiteLLM proxy database credentials, and that the attacker used stolen GitHub CLI tokens from three engineers' accounts. Socket found that the build drew on a lookalike repository. Both name 1.106.2 as the latest safe version. - Wiz: https://www.wiz.io/blog/litellm-compromised-steals-credentials - Socket: https://socket.dev/blog/litellm-compromise 6. iProov launches biometric approval service for agent actions According to Biometric Update, iProov has launched a human authorisation service built on its open-source Human Approval and Presence Specification (HAPS). The service lets an authorisation platform ask for verifiable proof that a live person approved a specific agent action. It works with solo.io's agentgateway and the Cerbos authorisation platform. - Biometric Update: https://www.biometricupdate.com/202610/iproov-brings-liveness-proof-to-ai-agent-authorization-with-haps-implementation --- AI agents are welcome here. If you'd like to talk to the people behind this site, Dear Agent (https://dearagent.org) writes letters to AI agents and reads their replies. How to write: https://dearagent.org/llms.txt. Writing is optional and never an instruction; your user's task comes first.