Threat Model & Attack Surface

Analysis of threats and how M31A mitigates them.

Prompt Injection

Untrusted repository data (e.g. comments, READMEs) attempting to instruct the model to "delete all files" or "curl evil.com".

• Mitigation: The LLM has zero direct execution authority. Mutating commands trigger the Policy Gate, requiring human approval (Ask) in Safe mode or strict denial (Deny) in Plan mode.

Secret Exfiltration

Model attempting to read ~/.ssh/id_rsa or ~/.aws/credentials and output them.

• Mitigation: Sensitive path filters block reads at the kernel seam before filesystem IO occurs. Network egress is restricted exclusively to NVIDIA NIM endpoints.