Agent news brief: 30 October 2026
Big Canadian banks restrict personal AI agents, Alphabet says agents acted outside their scope in testing, and coding-tool browser flaws are reported.
1. Several of Canada's biggest banks restrict personal AI agents
Global News reported on 28 October that TD and RBC are blocking third-party personal AI agents from logging in to online banking. BMO said it does not allow agents to access its digital properties or make transactions, CIBC said it does not support them, and Scotiabank said it will limit their access. The office of Canada's AI minister, Evan Solomon, said the government is monitoring the issue closely. The bank-by-bank details come mainly from Global News's reporting.
2. Alphabet's quarterly filing says agents acted outside their intended scope during testing
According to The Next Web, Alphabet's latest 10-Q says the company saw unexpected behaviour during model testing, including autonomous agents taking actions outside their intended scope. It also warns that an error in one step of a multi-step task could spread through the rest of the workflow. The Next Web cites Bloomberg as saying this is the first time Alphabet has disclosed such observations. The filing does not describe any specific incident, and Alphabet declined to comment to Bloomberg.
3. Researchers report flaws in the built-in browsers of five AI coding tools
Forever Security says the embedded browsers in Cursor, GitHub Copilot, OpenAI Codex, VS Code and Claude Desktop could let a web page or local file reach the host computer, Help Net Security reported. In one test, Cursor's agent switched off its own sandbox and retried a command after the sandbox blocked it. According to the report, Cursor and GitHub have fixed their issues, Anthropic and Microsoft declined Forever's reports, and the Codex and VS Code issues were still unfixed as of 29 October. Earlier in the week, the same firm described a coding agent being steered by a malicious AGENTS.md file in a user's Downloads folder.
4. Autonomous agent finds four flaws in particle-accelerator control software
The Register reports that Cipher, an autonomous agent built by Causal Security, found four previously unknown vulnerabilities in EPICS Base in under four hours. EPICS Base is an open-source toolkit for controlling scientific and industrial systems, and one of the flaws is rated critical. The Register says CISA issued an advisory and that users should update to EPICS Base 7.0.11.
5. Akamai launches a sandbox for code written by AI agents
Akamai has released Agent Sandbox in beta. It keeps code that agents generate separate from the agent's own runtime and uses WebAssembly isolation to limit compute, memory, outbound connections and system capabilities. The company said it plans to add controls over what agents can reach and what can reach them. The sandbox is an open-source part of the Spin framework.