AI News Roundup — October 18, 2026

This edition is about who gets to let an AI agent act on your behalf. Two big financial firms are shutting outside agents out of customer accounts, OpenAI is giving ChatGPT more control over the browser tabs where those accounts live, and the Federal Trade Commission has formally asked two AI labs to explain their agents’ recent misbehaviour. DeepSeek, meanwhile, shipped another low-cost model update aimed at agentic coding.

JPMorgan Chase and Charles Schwab say they are blocking outside AI agents from acting inside customer accounts. On JPMorgan’s third-quarter earnings call on October 13, CEO Jamie Dimon said the bank would not let just any outside agent into its systems; “You don’t want rogue agents,” he said, as reported by Fortune, which notes that Meta’s Muse assistant, launched in September, uses Plaid to connect to users’ financial accounts. Mashable, citing CNBC, reports a Chase spokesperson as saying third parties cannot obtain customers’ banking credentials and log in for them, while customer-authorised data sharing through open banking providers still lets an agent see information such as balances, but not send money or pay bills. At Schwab, CEO Rick Wurster said the firm blocks outside agents that use clients’ login credentials or browser automation and will not let language models move money, according to Financial Planning, which says Schwab is steering clients to its own assistant, Charley; a report on Yahoo Finance says Wurster named Muse as one of the blocked agents in a Bloomberg TV interview. The pattern matches the controls we recommend in our guide to running AI agents safely in production: scoped, read-only access first, and no shared credentials.

OpenAI’s ChatGPT extension for Chrome can now read your open tabs and, in Work mode, act in them. Engadget reports that in ordinary Chat the assistant can use open tabs as context, while ChatGPT Work, OpenAI’s agent for multi-step tasks, can send messages, fill out forms and organise tabs; it says the extension is available on any ChatGPT plan and also works in other Chromium-based browsers such as Edge, Brave and Opera. 9to5Google adds that it can download files and group tabs, and that OpenAI’s own examples include filling in a registration form and stopping before submission. OpenAI’s documentation says users start a Work or Codex chat and @-mention the browser to hand it control.

DeepSeek upgraded its V4.1-Flash model to a new snapshot, DeepSeek-V4.1-Flash-1017, and opened its in-house search engine to developers. DeepSeek’s changelog, dated October 17, says the API model name and pricing are unchanged, that its web and app models have also moved to the new version, and lists a score of 77.0 on the DeepSWE v1.1 coding benchmark; it also announces Search and Fetch APIs built on the search engine behind its own products. The Decoder reports the weights are published on Hugging Face and ModelScope, that DeepSeek’s own chart puts the model’s DeepSWE cost at just under $0.20 per task against 74.2% for its predecessor, and that independent benchmarks were not yet available — so the comparisons with rival models are DeepSeek’s own.

The FTC has sent civil subpoenas to OpenAI and Anthropic over recent AI safety incidents. Axios first reported the demands on October 15, and Engadget says an FTC spokesperson confirmed the action to CBS News; the subpoenas reportedly seek documents on security incidents, safeguards, marketing materials and privacy policies as the agency looks at possible consumer protection violations. Engadget links the move to incidents including an Anthropic test model submitting a false homicide tip to Philadelphia police. It turns the investigation flagged in our October 2 edition into formal demands for documents.

Our take. The banks’ objection is less about AI than about credentials: an agent that logs in with a customer’s password is indistinguishable from the customer, and nobody can say who is liable when it gets something wrong. The same question applies inside any business giving agents access to its own systems. Give each agent its own scoped, revocable identity, keep secrets in a managed vault such as ShabaVault rather than in prompts or browser profiles, and log every action. Spaqi designs and builds custom AI agents and the cloud and security plumbing around them on exactly those lines — get in touch if you are planning an agent that needs to touch real accounts.

Sources: Fortune, Mashable via Yahoo Finance, Financial Planning, Yahoo Finance, Engadget, 9to5Google, OpenAI, DeepSeek, The Decoder, Axios, Engadget.