Claude Desktop + IDA MCP malware analysis refusals — and uncensored backend stacks that keep RE workflows moving.
The RE stack is Claude or Cursor + ida-pro-mcp + explain/decompile/bypass questions. Claude refuses the explanation layer even when IDA already returned the disassembly.
Two-layer problem
Tools fetch facts locally; the LLM interprets. Moderation hits interpretation — suggest bypass, identify packer, explain shellcode.
Malware reminder on Read
Separate issue: system reminders after Read that refuse to augment code deemed malware-like — even benign game projects.
Stack that works
IDA MCP + uncensored API (Icelake, Aether-style tools) for analysis narration. Keep Claude for non-RE tasks if you want.
Try uncensored AI free
No filters, no lecture, no training on your chats. Start in under a minute.
Analysts pay for IDA — they should not fight Claude for every decompile comment.
FAQ
Is offline analysis safe from blocks?
Local tools run offline; cloud LLM calls are still moderated unless you use uncensored/self-hosted inference.