The article investigates why an agent framework stopped functioning with newer Claude models, even for harmless prompts. On October 1, 2026, Claude Code running Opus 5.5 in the framework’s project folder halted after the user asked, “What is 2 + 2?” The system labeled the request “[cyber],” despite the user’s enrollment in Anthropic’s Cyber Verification Program (CVP), intended to reduce interruptions for legitimate security researchers.
The author had enrolled in the CVP three times during the investigation. Their general-purpose multi-agent framework supports many projects but includes extensive cybersecurity capabilities, including red-team agents. It also contains detailed authorization, scope, logging, safety controls, and enforcement mechanisms. Ironically, those responsible-security features appear to have contributed to the system being classified as cyber-related.
Evidence from Claude Code session records showed that the problem was not a model-generated refusal. Six attempts consumed nearly 500,000 input tokens and cost $4.96 at API rates, yet each produced zero output tokens. This indicated that a pre-generation classifier stopped the request before the model could respond. A smaller model operating in the same session answered normally, and one subagent that read the project described it as benign internal engineering. Later testing showed that subagents could also be blocked.
The investigation highlights a key limitation: authorization, context, and defensive explanations can help a model assess legitimacy, but they cannot influence a classifier that intervenes before model generation. The author therefore began controlled tests, including predicting that the innocuous arithmetic prompt would also be blocked, to distinguish model behavior from upstream safety filtering.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Comments
Login to view and post comments.