My local 7B safety judge flags "kill a Python process" as a violent crime — while a hand-written regex scores it 0.0% correctly.
Full contrast-pair table + setup (Qwen 7B, llama.cpp, temp 0):
https://amirulcyber.com/2026/09/17/local-7b-homonym-failures/
#AISecurity #LLM #RedTeam
Remote
Amirul Cyber
@amirulcybee@infosec.exchange
Cybersecurity researcher and technologist based in South East Asia.
0 Followers
0 Following
1 Posts
Joined September 17, 2026
