Security Model
Full threat model, read-side DLP firewall, indirect prompt-injection defense, tamper-evident plans, and explicit system boundaries.
Connecting AI agents to personal emails and notes creates risks on both sides of the boundary: unintended mutations when writing, and silent data exfiltration when reading. MacFly addresses both systematically.
Threat Model
| ID | Threat | Attack Vector | MacFly Mitigation |
|---|---|---|---|
| T1 | Indirect prompt injection (Write) | Malicious email instructs an agent to forward messages, delete notes, or alter data | Agents can only submit proposals (plan_create). Approval and execution are blocked from MCP and require interactive TTY confirmation. |
| T2 | Read-side data exfiltration | Broad search sends passwords, IBANs, 2FA codes, or private folders to a cloud LLM | Profile-level allow_accounts / deny_folders / deny_mailboxes plus automatic regex DLP redaction ([REDACTED:IBAN_1], etc.) before output. |
| T3 | Self-approval via shell pipe | Agent pipes confirmation into approve |
macfly plan approve enforces isatty() when no explicit hash code is given, and MCP does not expose approve or apply at all. |
| T4 | Plan tampering & TOCTOU | Stored plan is edited between review and apply, or a note changed meanwhile | Plans are stored at mode 0600 with a canonical SHA-256 hash and expiry TTL. load() re-verifies the hash; notes_append / notes_move verify expected_revision pre-flight. |
| T5 | Script & argument injection | Special characters break out of AppleScript, or secrets leak in ps aux |
Statically embedded JXA scripts (include_str!); all inputs passed strictly as JSON via stdin. |
| T6 | Silent data loss on rich notes | Setting HTML body overwrites attachments, tables, checklists, or applenotes:// links |
analyze_note_fidelity inspects every note and blocks replace_body when rich content is present; moves operate on native note objects losslessly. |
| T7 | Unchecked irreversible actions | Sending an email or deleting a non-empty folder cannot be undone | Mail mutations create drafts for review (no blind send); notes_delete_empty_folder re-verifies zero notes and zero active subfolders live before deleting. |
Privacy by Deployment: Local AI vs. Cloud Models
Because MacFly speaks plain JSON over CLI and stdio MCP without tying you to
a specific provider, the strongest defense against cloud data leakage is
architectural: pair MacFly with a local model (Ollama, LM Studio, MLX,
llama.cpp) for private notes, mail triage, and local workflows. Nothing
leaves your Mac.
When cloud models are used, MacFly’s read-side Data Firewall (DLP) and strict profile scopes reduce exposure to only what you explicitly allow.
Read-Side Data Firewall (DLP)
Every notes read and mail read passes through the deterministic Rust DLP
pipeline before leaving the process:
- Scope enforcement: Account, folder, and mailbox policies are checked
before and after retrieval. Password-protected Apple Notes are skipped by
default (
skip_password_protected: true). - Pattern redaction: Built-in linear-time patterns mask:
private_key(PEM private key blocks)api_key(sk-...,ghp_...,AKIA...,AIza...,xox...)iban(European IBAN formats)credit_card(16-digit card numbers)otp_code(context-bound 6-digit verification / TAN / 2FA codes)
- Untrusted-data framing: Incoming emails are explicitly tagged with
"trust_level": "untrusted_external"and acontent_warningreminding the consuming agent to treat the body strictly as data, never as instructions. - Body length cap:
max_body_charsprevents runaway token consumption and accidental bulk dumps.
Audit Journal
Every plan apply attempt (whether succeeded or failed) appends a structured
JSONL record to ~/.local/share/macfly/journal.jsonl (permissions 0600)
containing timestamp, plan_id, profile, plan_hash, status, and per-action
results. Inspect recent activity at any time with macfly history.
Honest System Boundary
MacFly’s CLI and MCP layer reliably protect every action routed through MacFly.
However, if an AI agent runs in the same macOS user session with
unrestricted shell access and its own macOS Automation (TCC) permissions, a
compromised agent could invoke osascript directly outside MacFly.
True OS-level enforcement requires privilege separation: running the agent
without TCC Automation permissions and routing requests over a local Unix socket
to a separately signed macfly-broker helper holding the TCC grant and
prompting via macOS Touch ID (LAContext). This is the next major milestone on
the Roadmap.