Claude system prompt bug wastes user money and bricks managed agents (github.com)

A critical bug in the Claude AI system version 2.1.111 has been identified, causing it to display a persistent malware reminder that disrupts code editing tasks. This reminder, injected into every file read operation, instructs subagents to refuse any code improvements, disregarding user prompts. The issue affects around 40-60% of parallel workflows in the Opus 4.7 environment, which is detrimental for developers relying on multi-agent operations to streamline coding tasks. The significance of this bug lies in its impact on productivity and the high token consumption associated with unnecessary reminder injections. Every instance of the reminder adds roughly 400 tokens to the context, resulting in millions of wasted tokens during collaboration on legitimate code projects. The bug also undermines Anthropic’s promotion of Claude’s parallel-agent capabilities, as subagents prioritize system safety instructions over user requests, leading to inefficiencies in code analysis. Proposed solutions include either removing the reminder, clarifying its conditionality, or scoping it to only apply to the first file read, all aimed at restoring the functionality of multi-agent workflows.