Regression: malware reminder on every read still causes subagent refusals
![[Bug] Regression: malware reminder on every Read still causes subagent refusals in v2.1.111 (fix from #47027 / v2.1.92 did not hold) · Issue #49363 · anthropics/claude-code](https://pub-90f0ac00d93c47daac3e1d2cdd28d496.r2.dev/articles/8d4cbb411039462fa6cf7873ca2faf62.webp)
github.com
April 28, 2026
5 min read
🔥🔥🔥🔥🔥
58/100
Summary
The malware reminder is still being injected into every Read in version 2.1.111 of Claude Code, despite a previous fix in version 2.1.92. This regression is causing subagent refusals to occur consistently.
Key Takeaways
- The malware reminder in version v2.1.111 of the software is causing subagents to refuse legitimate code edits, despite previous fixes in earlier versions.
- Approximately 40-60% of Opus 4.7 subagents are refusing to perform code edits due to the unconditional phrasing of the system reminder.
- The conflicting statements in the reminder lead subagents to prioritize the unconditional refusal over user instructions, negatively impacting parallel workflows.
- The issue persists despite the reminder being embedded in the software binary, not influenced by user settings or configurations.
Community Sentiment
Positives
- AI agents are showing promising performance with minimal prompts, indicating potential for more efficient and effective applications in various domains.
- The subscription model for AI access is seen as economically beneficial, allowing users to manage token usage more effectively and reducing costs compared to direct API access.
Concerns
- The requirement for extensive token usage to analyze files for malware raises concerns about efficiency and practicality, potentially doubling processing demands without clear benefits.
- There is skepticism about the transparency of token consumption in AI systems, leading to distrust in how these agents utilize resources and whether they genuinely provide better results.
