grpo-rl-training
Warn
Audited by Socket on Sep 9, 2026
1 alert found:
SecuritySecurityexamples/reward_functions_library.py
MEDIUMSecurityMEDIUM
examples/reward_functions_library.py
Overall risk is dominated by unsandboxed `exec()` of Python extracted from untrusted LLM content. While the module is intended for evaluation/scoring, it effectively acts as a code execution engine during reward computation, enabling arbitrary code execution (and thus potential data theft, sabotage, or environment compromise) if attacker-influenced model outputs are possible. No explicit hardcoded credentials or direct network/exfiltration endpoints are visible in this fragment; however, the execution primitive makes malicious behavior feasible. Additional sinks may exist in omitted referenced helpers, but the `exec()` path alone is a critical red flag.
Confidence: 86%Severity: 90%
Audit Metadata