Claude refused to finish patching my servers. My local LLM took over

On October 21 and 22, I spent two days patching our servers due to critical security vulnerabilities in PHP. I used Hermes, my AI agent, to assist me. However, the cloud model ran into issues and stopped responding during the process. My local model, Qwen3.8-Flash-Next, took over and managed to finish the work.

The patching process involved several phases: updating PHP on all servers, cleaning up and restoring compromised WordPress sites, and creating an inventory of all plugins on every site to look up known vulnerabilities. However, the local model was unable to continue the work due to content filter restrictions triggered by the cloud model's safety classifiers.

Claude Opus 5.5, the cloud model, refused to respond to further requests, citing restrictions on cyber content. The refusal was invisible to the model, and the agent could not inform me about the issue. After a few retries, the session still resulted in the same refusal, even with different commands like continue, try it another way, and try again differently.

The final refusal message from Claude Opus 5.5 mentioned that the request triggered restrictions on violative cyber content. This category covers requests that could enable cyber harm and even benign cybersecurity work can also trigger this category. The context window at the time of the refusal included access log lines, CVEs, plugin version diffs, and forensics on real web shells and backdoor admin accounts.

In conclusion, the cloud model refused to finish patching my servers due to content filter restrictions triggered by its safety classifiers. My local model, Qwen3.8-Flash-Next, stepped in and completed the work successfully.