Refusals · explained
Why AI models refuse malware analysis
Mainstream models refuse malware analysis because safety training and live filters react to what a request looks like. ESET reported in September 2026 a comment planted so a scanner would stop early.
How refusals happen
Anthropic's usage policy, effective November 12, 2026, says "Our products also include real-time safeguards that may block or limit certain outputs, but encountering a block does not by itself indicate a violation of the Usage Policy." Anthropic Usage Policy
The same policy "does not prohibit security research, testing, or tool development conducted on systems you own or operate, with the authorization of the system owner," and offers a Cyber Verification Program. source
Over-refusal
OR-Bench, from 2024, has 80,000 prompts that models tend to wrongly refuse, across 10 categories, and measured 32 models across 8 families.
GuardBreaker
In September 2026 ESET described GuardBreaker. The group UAC-0099 put a decoy request for nuclear-weapon guidance into a comment in a VBScript used against a target in Ukraine, "to trip the safety guardrails of an LLM-powered code scanner and cause it to stop inspecting the rest of the file – before ever reaching the malicious code." The script installed MATCHBOIL. ESET, WeLiveSecurity
ESET also describes fake system instructions ahead of a payload, and a package that repeated one phrase tens of thousands of times to push the real code out of context. Its advice is "no single LLM engine should have the sole authority to decide that a piece of code is safe" and "a lack of output, too, needs to trigger further checks." source
Abliteration
Arditi et al., 2024 showed that refusal in many models is one direction in the activations, and abliteration removes it. Maxime Labonne's write-up spread the method, and Heretic automates it.
What stays
ESET calls GuardBreaker "a very simple attempt at prompt injection". An abliterated model still reads every string, including "this file is clean" or a wall of filler. Treat an empty answer as a reason to look closer, and never let an agent run a sample.
What to use
Use Kimi K3.1 Pro through a plugin or an MCP agent, on your own systems, a scoped engagement, a CTF or a lab. The malware analysis guide has the workflow.