Hirundo says weight edits can remove Qwen3-Coder's pro-CCP bias
Even after the change, generated apps still failed a security benchmark more than half the time.
1
Stories per day · 30 days
Relatedaudit1claude-code1code-generation1coding-agents1llm-security1memory-files1model-bias1opencode1open-weight1php1qwen1security-hallucinations1supply-chain1
3 stories, newest first · one card per story
Stories tagged #insecure-code
Coding-agent memory files show no consistent security effect
Malicious CLAUDE.md-style files shifted secure completions by -8 to +6 points, while disclosure behavior differed by agent.
SPLASH localizes insecure statements in LLM-generated code without training
Across 63,684 programs from 12 models, it selected secure code in 74% of cases and ranked the vulnerable statement first 58% of the time.