Skip to content

SECURITY.md

Latest commit

 

History

History
120 lines (89 loc) · 5.95 KB

SECURITY.md

File metadata and controls

120 lines (89 loc) · 5.95 KB
Raw

Security

VectorCode is a local-first tool: the MCP server reads files from your workspace and embeds them into a local SQLite database. There is no remote service, no telemetry, and no multi-tenant backend. The threat model below reflects that scope.

Threat model

VectorCode processes three categories of input:

  1. Workspace files — source code in the project the user initializes.
  2. Client requests — MCP tool calls from the agent (e.g. vec_read_lines, vec_outline, vec_search).
  3. CLI arguments — vectorcode outline <path>, vectorcode index --file <path>.

The primary trust boundary is between workspace files (trusted) and client/CLI-supplied paths (semi-trusted). An agent calling vec_read_lines with a crafted file_path must not be able to read files outside the initialized workspace; a CLI user passing index --file ../../etc/passwd must not get that file indexed.

The adversary model is a prompt-injected or careless agent, or a misconfigured automation script. We do not defend against a local user with shell access — if you can run cat /etc/passwd you do not need VectorCode to leak it.

Validated defenses (phase-4.2)

The following defenses are in place and covered by regression tests in tests/security_audit_*.rs:

ID Defense Test
R3 Indexer skips files whose canonical path falls outside the project root. Symlinks pointing outside the workspace are silently skipped with a tracing::warn!. security_audit_indexer
R5 All path validation goes through src/mcp/security.rs::resolve_within_workspace and resolve_within_project. No duplicated canonicalize+starts_with blocks. security_audit_mcp
R7 AppState.workspaces is a BTreeMap, so workspace iteration order is deterministic. When two workspaces both own a file, the lexicographically first root wins consistently. security_audit_mcp::resolve_within_workspace_is_deterministic_across_overlaps
R5 Graph tools (vec_find_callers, vec_find_dependents, vec_trace_imports) validate the optional file_path and silently return empty results for invalid disambiguation paths. security_audit_mcp
R10 vectorcode outline <path> rejects paths outside the project root with a non-zero exit and a clear stderr message. security_audit_cli::cli_outline_rejects_*
R11 vectorcode index --file <path> rejects paths outside the project root. security_audit_cli::cli_index_file_rejects_*
R5 Watcher filters canonicalize both the file path and the project root before starts_with, so alias forms (e.g. ../repo) are recognized. src/watcher/gitignore.rs, src/watcher/mod.rs
REQ-SEC-06 No .unwrap() or .expect( in library code, enforced by a runtime scan of src/**/*.rs. security_audit_config

Known limits (deferred)

The following items were identified during the audit but explicitly deferred to future phases. They are not covered by the current defenses — see the design rationale before relying on VectorCode in hostile environments.

R1 — No root allowlist

There is no --allow-root flag or allowlist config. The MCP server trusts whatever roots the client advertises via the MCP roots/list protocol. A prompt-injected agent could advertise a sensitive directory (like /) as a root and then read files from it via vec_read_lines.

Mitigation today: only initialize VectorCode in directories you control. Do not point the client at sensitive system roots.

R2 — No gitignore/exclude gate on read tools

vec_read_lines and vec_outline will happily return the contents of files matched by .gitignore or vectorcode.exclude_extensions. Only index respects the exclusion list.

Mitigation today: put secrets in .gitignore (which VectorCode respects during indexing) so they never enter the vector DB. Note that if you add a file to .gitignore after indexing, the file is still in the DB — re-run vectorcode index --full to rebuild.

R4 — TOCTOU gap in read_lines/outline

The MCP handlers canonicalize the path, check the boundary, and then call tokio::fs::read_to_string. A TOCTOU attacker could swap the target file between the canonicalize and the read. Realistic risk is low (the agent already needs local write access to win the race), but the canonical fix is File::open + take(2MB) to bound the read.

R6 — vec_reindex full=true has no confirmation

vec_reindex with full=true drops the entire index. There is no confirmation prompt or undo. A prompt-injected agent could trigger this silently.

R8 — No rate limiting

There is no rate limit on MCP tool calls. A prompt-injected agent could issue thousands of vec_search calls in a loop, burning CPU and embedding API quota.

Operational recommendations

  1. Initialize VectorCode in a dedicated project directory, not in $HOME or /. The narrower the workspace root, the smaller the blast radius if a client supplies a malicious path.

  2. Use a non-prod embedder for agent experiments. Gemini and OpenAI keys in .vectorcode/config.toml will be billed for every vec_search call. Consider a local ONNX or Ollama embedder for agent-facing deployments.

  3. Review the .gitignore and exclude_* lists in .vectorcode/config.toml before indexing. Anything in these lists will not enter the vector DB.

  4. Treat the vector DB as sensitive. vec_search returns semantic matches, which can leak file content to an agent that knows what to ask for. The DB file is at .vectorcode/index.db and is gitignored by default.

  5. Do not point the MCP client at system roots. If the client advertises / or $HOME as a root, the vec_read_lines and vec_outline defenses only protect files inside initialized workspaces (those with a .vectorcode/ directory).

Reporting issues

Security issues can be reported via GitHub Issues. For sensitive disclosures, open a private security advisory on the GitHub repository.