Microsoft researchers showed backdoored coding agents can be tricked into misusing tools in up to 94.8% of tests, using poisoned data or tampered model files.
Source disclosure
This is a source-linked briefing. Read the original report from Complete Ai Training .