relift

Relift — Roadmap

Phases, not dates. A phase is done when its exit criteria are met, not when the calendar says so.

v0.1 — Walking skeleton (shipped)

Scope: engine, catalogue, scope gate, stdlib modules, AI abstraction, reporting.

  • Packaging, relift entry point, python -m relift fallback
  • Catalogue with tiers, MITRE hints, implementation status
  • Scope gate: allow/deny, CIDR, wildcard, expiry, risk ceiling, wide-range guard
  • Dry-run default; --execute required
  • Stdlib modules across recon / tls / web
  • External wrappers for nmap, nuclei, nikto with fixed argv templates
  • Local, cloud and offline provider matrix behind one interface
  • Audit JSONL + per-run artefact + text/markdown/json rendering

Exit: list renders; a lab host produces ≥6 executions; --no-ai findings are identical to an AI run.

v0.2 — Depth (this tree, 0.2.0)

Theme: make one run worth reading twice, and let it read the things that hold the data.

  • Database auditing: real wire-protocol fingerprint and security-posture checks for PostgreSQL, MySQL/MariaDB, MSSQL, Redis, MongoDB, Elasticsearch and CouchDB (db.fingerprint, db.security_config)
  • Held-credential verification: one attempt per operator-supplied credential, plus privilege level per engine (db.credential_verify, lab only)
  • Bounded SQL injection detection: error, boolean and time probes against one named parameter, 8-request cap, no extraction (web.sql_injection_check)
  • Local privilege-escalation surface audit: SUID/SGID, capabilities, sudo rules, container sockets, listening sockets, kernel advisory pass (system.host_audit, read-only, Linux)
  • Attack threads: findings correlated into multi-stage paths with an explicit break point (analysis.attack_chains, relift chains)
  • Target families: db://, ssh://, local:// targets, plans filtered by family so a datastore is never sent an HTTP template scan
  • SSH algorithm audit implemented natively — the ssh-audit dependency is gone
  • Desktop application: loopback-only API, native window via pywebview, browser fallback, relift desktop / relift-desktop
  • Terminal, Markdown and JSON reports carry the threads

Exit: a lab datastore yields a critical exposure finding and a thread; a held credential is verified once and graded; the desktop app runs a scan through the same scope gate as the CLI.

v0.3 — Breadth

Theme: internal network, not just one URL.

  • Multi-target runs: --targets FILE with one scope check per host, one run artefact per host, one combined report
  • Passive subdomain inventory from certificate transparency logs (recon.passive_subdomains; still no DNS brute-force)
  • Directory enumeration policy for web.dir_bruteforce: a wordlist budget, 404 calibration and a rate ceiling, or it stays a spec
  • Scheduled assessments: a documented cron/systemd-timer recipe plus a --fail-on <severity> exit code for CI
  • relift diff <runA> <runB> — new / fixed / changed findings and threads
  • HTML report, single self-contained file, opens offline
  • tls.cipher_preference once sslscan output shape is pinned
  • Bounded parallel execution sharing one global rate budget
  • Test suite: scope matcher, target parsing, runner, chains rules, each module against a local fixture server

Exit: a /24 sweep stays inside the budget, every host gets its own plan and artefact, and two sweeps a week apart produce a diff a human can act on.

v0.4 — Analyst quality

Theme: the AI should be interesting, not merely present.

  • JSON-mode structured triage with schema validation and repair
  • Prompt-injection test suite (hostile banner → analyst must ignore it)
  • Per-provider quality harness on a fixed finding corpus
  • Redaction pass on by default for cloud providers
  • Chain-aware narration: the model is given the threads and their break points, not just the finding list

Exit: the injection suite passes for all providers; a model that fails schema validation falls back instead of guessing.

v0.5 — Hardening

  • Fault injection: module raises, DNS fails mid-run, provider dies
  • Windows path/encoding pass for the host audit (currently Linux-only by design)
  • Security review of the scope gate by someone who did not write it
  • Security review of the desktop API surface (token handling, Origin check, job lifetime)
  • Fuzz the protocol parsers in dbproto.py with malformed frames

Exit: the gate cannot be bypassed by any documented or documented-adjacent route, and a malformed database frame produces a finding, not a traceback.

v1.0 — Release decision

  • External review of 03-RULES-OF-ENGAGEMENT.md
  • Relicense (Apache-2.0 vs AGPL-3.0 — Jay's call)
  • CONTRIBUTING with the tier-D exclusion stated as policy
  • Public docs site or a very good README

Exit: a stranger clones, reads for ten minutes, and cannot accidentally scan a host they do not own.

Explicitly not on the roadmap

Denial of service · exploit delivery · credential brute-force and spraying · default-credential checking · data extraction or dumping · persistence · lateral movement · exfiltration. These are excluded by policy, not deferred by priority. See 03-RULES-OF-ENGAGEMENT.md §7.

Idea parking lot

Things worth writing down, not worth promising:

  • A --json output shape stable enough to feed a small dashboard
  • Signed scope files (ed25519) so a scope cannot be edited after sign-off
  • Per-module request budgets instead of one global rate
  • A "pre-flight" mode that reports what would be skipped and why, before --execute
  • Comparison against a baseline posture file ("this is what my lab looked like last month")
  • Chain rules contributed as data (a small predicate language) instead of Python
← AI ProvidersCLI Reference →