Skip to content

Security: GiulioDER/RE-call

Security

SECURITY.md

Security Policy

RE-call protects the memory retrieval boundary. It does not redact, encrypt, or classify the corpus. The main asset is the text you index.

Supported Versions

Pre-1.0, only the current 0.x line gets fixes. There is no LTS branch.

Version Supported
Current 0.x release Yes
Older 0.x releases Best effort only

Report a Vulnerability

Open a private security advisory on GitHub, or contact the maintainer through the repository owner profile. Do not paste private corpus text, credentials, tokens, DSNs, or API keys into a public issue.

Useful reports include:

Area Report when
Tenant isolation One tenant can read another tenant's rows through the library, MCP server, or integrations.
Index confinement recall_index can read files outside the configured index root or outside the allowed glob.
Credential handling A DSN, token, or API key is logged or returned in an exception, tool response, or result.
Cloud egress A cloud embedder is reached without the caller explicitly selecting it.
Trust refusal Strict mode returns corpus text when calibration or generation identity is not certified.

Expected behavior that is not a vulnerability:

Behavior Why
Retrieved chunks contain the original memory text Returning memory verbatim is the purpose of the library.
A tenant principal can read all chunks for its own tenant Isolation is tenant-level, not per chunk.
Voyage embeddings send chunk text to Voyage That is the explicit cloud embedder path. Use the local FastEmbed path for private corpora.
Local Docker credentials are recall:recall They are for local development only; production DSNs must use different credentials.

Deployment Rules

  • Run application traffic with an unprivileged serving role.
  • Keep the migration DSN out of the MCP process.
  • Set RECALL_INDEX_ROOT to a directory that contains only indexable corpus files.
  • Keep token files and .env files outside the index root.
  • Use local embeddings for sensitive corpora unless cloud egress has been approved.
  • Treat calibration artifacts as tenant data, because they can reveal corpus topics and evaluation intent.

Detailed threat model: docs/SECURITY_MODEL.md.

There aren't any published security advisories