Users Continue to Trigger Abnormal Outputs and Hallucinations in Claude Opus 5

Hydiin · reddit · 2026-07-31

Recently, multiple users have reported frequently encountering abnormal outputs and hallucinations when using Claude Opus 5.

One user noted that by experimenting with different chat windows and specific types of messages, it is still possible to bypass the model's safety mechanisms and trigger erratic responses. It appears Anthropic is internally working on patching these prompt injection and guardrail bypass vulnerabilities.

Related event: Specific Prompts Trigger Abnormal Completion and Jailbreak in Claude Opus(19 posts)→

Original post →

More from Models

Models channel →