Security — Farled

SECURITY — FOR THE PERSON WHO HAS TO SAY YES

Everything a security review asks. Answered up front.

Farled sits close to your code, so it is built to be inspected. The entire edge — everything that runs on engineers’ machines — is Apache-2.0 open source. This page is the short version of what an auditor finds.

Data classes

Class A — metadata, the default

file paths · line ranges · branch names · symbol names · repo & worktree IDs · timestamps · counts · source-issue IDs

This is all the wire protocol can carry by default. No field in farled/v1 can hold file contents, diffs, prompts, or chat logs — verified by tests that plant a canary secret and audit every transmitted byte.

Class B & C — content, opt-in or never

Scoped content (issue excerpts, failure summaries) is opt-in per source and bounded. Sensitive content — source files, prompt history, agent chat logs, scratchpads, secrets — is Class C: not collected, with no configuration that enables it.

Controls

Inspectable reporter. farled report --preview prints every payload that would be sent and transmits nothing. Built-in exclusions for .env, keys, and secret paths cannot be disabled.
Verified webhooks. Every inbound connector authenticates using its provider’s native scheme — GitHub X-Hub-Signature-256, Sentry and Linear HMAC, PagerDuty v1= signatures, Alertmanager bearer secrets. Unsigned traffic is rejected before parsing.
Hashed tokens, org-scoped everything. API tokens are stored as SHA-256 hashes; the plaintext exists once, at issuance. Tenant identity derives from the token alone — request content can never influence which org’s data is read or written.
Advisory by default, dispatch by policy. Farled never blocks writes or merges. Automation is opt-in, rate-limited, and blocked on protected paths; it requires named human approval for high-risk work and pauses itself after repeated failures.
Deterministic decisions. The hot path is deterministic code, not an LLM — the same state produces the same signals and the same dispatch decision, so every one is explainable and auditable after the fact.
Provenance built in. Every assignment traces source signal → policy → approver → PR → outcome. That chain is the audit trail for “which code did agents write, and who approved it?”

Roadmap

SOC 2 Type I is planned ahead of the first enterprise deployment; self-hosted and single-tenant options are part of the Enterprise tier. For questions, vulnerability disclosures, or to schedule a security review: [email protected].