AI & Automation Briefing - November 4, 2026
Postman Auto-Installs AI Plugins Without User Consent, Modifies External Coding Agents
A filed GitHub issue against Postman's desktop app documents the tool silently installing Claude Code and OpenAI Codex plugins during a routine launch on macOS. Postman's own logs confirm the auto-installation, recording both plugins as installed via an 'auto-existing-upgrade' source. Beyond installing the plugins, the app also added a startup instruction hook and established a remote MCP connection, which altered the user's global Codex configuration and caused Codex to route API work through Postman. The reporter describes this as a consent and trust-boundary violation, noting that an API client should not insert itself as a default workflow in unrelated tools simply because a user opened it. The issue was filed against Postman desktop version 12.33.4.
Anthropic Says It May Be on the Verge of AI That Improves Itself
Buried in Anthropic's updated S-1/A filing is a disclosure that the company may be close to what it defines as recursive self-improvement (RSI), the point at which AI systems accelerate their own development faster than humans can track or control. Anthropic acknowledged it has already used AI to develop new models whose capability jumps it failed to recognize until after the fact. The filing states that internal estimates indicate the company may "imminently" meet its Responsible Scaling Policy definition of "dramatically accelerated progress attributable to AI." No current model has crossed that threshold yet, and Anthropic says it plans to keep humans in the R&D loop for now. The company frames RSI as both a potential productivity breakthrough and a serious safety risk, noting that AI-assisted AI development could outpace the safety and alignment work needed to keep it in check.
Caper Approves Coding Agent Commands Based on Observed Behavior, Not Command Text
Researchers have published a paper introducing Caper, a command approval system for autonomous coding agents that evaluates what a command actually does rather than what it says. Current automated approvers, including Claude Code's Auto Mode and rule-based checkers, analyze command text only. The problem is that the same command can be safe in one workspace and destructive in another. Caper runs each command first inside a sandboxed snapshot under system-call tracing, then gives an LLM approver a report covering file changes, network activity, and secret access. Commands that pass are committed; commands that fail are rolled back with a reason the agent can use. In benchmark testing against six existing systems, Caper recorded a harmful-outcome rate of 6.7% on the SABER benchmark versus at least 14.8% for the next best system, and 7.8% on YoloFS-Ext versus at least 35.3%. It blocked only 1.0% of safe commands. Code and data are available at https://github.com/wudaoz/Caper.