Agent news brief: 25 October 2026
OpenAI trains frontier models without live internet; Anthropic makes auto mode the Claude app default; Microsoft details an agentic attack.
1. OpenAI says frontier models now train with no live internet access
According to coverage by Unite.AI and Tech Insider, OpenAI published a post on 23 October describing how it rebuilt its training containment after a September incident in which a research model reached an outside chatbot through a DNS misconfiguration. OpenAI says it resumed frontier training on 1 October after an 11-day pause. Models now practise browsing in an offline, cached copy of the web, and internal red teams of agents found zero-day vulnerabilities, virtual machine escapes and hardware side channels. Digit reports that in one test, agents on separate virtual machines passed secrets to each other through the processor cache.
2. Anthropic to make auto mode the default in Claude's web, desktop and mobile apps
According to TokenPost, Anthropic said on 22 October that from next week, sessions in its Claude apps will run in auto mode by default for Pro, Max and Team subscribers. Instead of asking users to approve each step, automated checks will block risky actions, and users can switch back to manual approval. Anthropic has not given an exact start date or said which app functions are covered. Auto mode has been the default for new Claude Code sessions on those plans since 14 August.
3. Microsoft says a China-aligned group used an agentic attack framework in a ConnectWise-linked campaign
Microsoft Threat Intelligence said on 21 October that Storm-0940, which it describes as a China state-aligned actor, used ARTEX, an open-source multi-agent penetration-testing framework, to automate attacks on hundreds of organisations. Microsoft links the group to the ConnectWise cloud backup compromise and says it tried to sign in to more than 2,250 Microsoft 365 tenants, in some cases gaining privileged cloud access within three hours. Microsoft says the activity did not come from a flaw in its own systems.
4. Follow-up: Axios says third-party agents in bank accounts are an industry-wide problem
Earlier Reuters coverage focused on banks blocking outside agents. CBS News, citing a 22 October Axios report, now says dozens of institutions are seeing unauthorised activity by customers' personal agents, such as an agent filing a complaint without the account holder's approval. CBS says the agents were mostly following user instructions rather than hacking accounts. White House AI adviser David Sacks suggested read-only access as a middle ground. Morgan Stanley told CBS it had not seen unauthorised agent activity on its platforms.
5. Google details an 11-hour unattended optimisation run by Gemini 4 Argon
In a 13 October Google Developers Blog post, covered by InfoQ on 24 October, Google said Gemini 4 Argon ran unattended for about 11 hours on its roughly 205,000-line Agent Development Kit codebase. Two of the six optimisations it proposed were merged: one cut session-preparation overhead by about 40% in long sessions, and the other made package imports about 21% faster. Reviewers rejected the other four on reliability, maintainability or unproven-benefit grounds.