If you believe you have found a security vulnerability in Numba, please open a security advisory here: https://github.com/numba/numba/security/advisories/new

We will respond promptly to security issues, and if we do not believe it is a security issue, we may ask you to open a public issue instead.

Numba is a just-in-time compiler which can generate arbitrary LLVM IR, compile it, and load it immediately into the running process. As a result, the compiler should not be used with untrusted inputs. This includes both the function being compiled, but also arguments to the compiler decorators (@jit, @vectorize, etc) that control code generation. Additionally, Numba-compiled code, like C and FORTRAN, is not intended to be memory-safe.

Numba's on-disk function cache (see numba/core/caching.py and the caching docs) uses Python's pickle module to serialize and deserialize compiled artifacts. Deserializing data with pickle can execute arbitrary code, so a Numba cache directory must be treated with the same level of trust as the code you execute. Do not point NUMBA_CACHE_DIR at shared, world-writable, or otherwise untrusted locations, and do not use cache contents from untrusted sources.