Security and Privacy

Assets and threats

ThreatMitigation in codeProof
Prompt injection via repository text, model cards or MCP descriptionsContent is treated as data; the engine never executes or follows it; reports escape itaudit A16, fuzz renderer cases
Secret leakage (.env, keys, credentials).env content never read; .pem/.key/.ssh/.aws refused as evidence; URL credentials redactedA05, A07, A08
Path traversalpath.relative containment, workspace itself refusedA03
Symlink substitutionlstat rejection plus realpath containmentA04, A09
Oversized or hostile files2 MiB cap, 200 manifests, bounded parsing, fuzzedA06, A23, fuzz
Log tamperingSHA-256 hash chain (detects edits, deletes, reorders)A01, A02
Report tamperingCanonical SHA-256 sealA29
API abuseLoopback only, bearer token, GET only, no CORS, rate limitA19, A20
Adapter overwrite and host traversalPlan-then-copy, --force, strict host regexA27, A28
False confidenceUnknown is not pass; decisions always need a humanA10-A15, A17, A18
Supply chainZero dependencies; SHA-pinned CI actions; npm provenancenpm ls

Privacy

No network, telemetry, account or API key. Data kept: system metadata you enter, SHA-256 digests, relative evidence paths, audit events. Not kept: file content, .env values, prompts, personal data. You are responsible for what you type into records; do not register real customer data.

Residual risks

Reporting a vulnerability

Private GitHub security advisory. See SECURITY.md.