Highlights
Command Injection: 2 prior fixes. Scrutinize any change in this area.
npcpy/npc_compiler.py: most-fixed (4 issues). Treat as high-risk during review.
5 high-severity fixes in this history; regressions here are high-impact.
Recurring patterns
The bug types that recur here, drawn from past fixes, not open vulnerabilities.
Command Injection: Repeated vulnerabilities allowed arbitrary shell execution via unsafe subprocess invocations with shell=True and string interpolation within file and web search tools. Mitigation requires strict adherence to list-form command execution without shell interpretation.
Insecure Deserialization: Arbitrary code execution was possible when loading untrusted weights/checkpoints using PyTorch. Restricting deserialization to safe tensor data types is necessary to mitigate this threat.
Denial of Service: Dynamic code execution using the 'exec' built-in was vulnerable to process termination via unhandled SystemExit exceptions raised by remote users.