relift

Relift — Project Plan

Field Value
Project Relift
Company Pixelcity
Developer / owner Jay
Version 0.2.0 (internal)
Status Engine + database auditing + host audit + attack threads + desktop app
License Proprietary for now, open source later (see LICENSE)

1. Problem

Testing your own infrastructure against realistic attack techniques is normally a pile of shell history: one nmap here, a nuclei template there, a curl -I to check headers, a browser to look at the certificate. The output is scattered, the ordering is improvised, and nothing records what was tested and what was not.

Relift turns that pile into one command with one output. The catalogue is the product: it states exactly which techniques exist, how invasive each one is, and which of them have actually been run against a given target.

2. Goals

# Goal Measured by
G1 One CLI entry point for the whole assessment loop relift list → plan → scan → report end-to-end on a lab host
G2 The tool documents itself relift list shows every technique, its risk tier and status
G3 Deny-by-default execution Zero module runs without a matching, unexpired scope entry
G4 Optional AI, never required Full run with --no-ai, byte-identical findings to an AI-enabled run
G5 Swap AI providers in one flag Same engine output across Ollama / OpenAI / offline
G6 Reconstruct any past run Per-run JSON artefact + append-only audit log
G7 Extensible without forks A new technique is one file plus one catalogue entry

3. Non-goals (deliberate, permanent)

  • No exploitation of third-party systems. Ever. Not a feature.
  • No destructive capability in the shipped tree: no denial-of-service, no data tampering, no payload delivery, no persistence, no lateral movement automation.
  • No credential brute-force engine. The catalogue documents the technique class and the module stays unimplemented by design (see 02-ATTACK-CATALOG.md, tier D).
  • Not a hosted service / SaaS. Local CLI only.
  • Not a compliance scanner. It does not try to certify anything.

4. Users

  1. Jay — primary. Runs it against Pixelcity infrastructure and a homelab.
  2. Pixelcity engineers — run it before shipping a service, read the report.
  3. Future self / contributors — after open-sourcing, need the catalogue and the scope model to be obvious within ten minutes.

5. Workstreams

WS Name Scope Owner
WS-1 Core engine Catalogue, scope gate, runner, findings model, audit Jay
WS-2 Modules Passive → active techniques, stdlib implementations Jay
WS-3 External wrappers nmap / nuclei / whatweb / nikto, fixed arg templates Jay
WS-4 AI layer Provider abstraction, local + cloud adapters, rule engine fallback Jay
WS-5 Reporting Terminal, Markdown, JSON; AI narrative Jay
WS-6 Safety & legal Scope schema, authorisation gate, excluded-class policy, docs Jay
WS-7 Packaging & release pyproject, relift entry point, versioning, changelog Jay

6. Milestones

M1 — Walking skeleton (done in this scaffold)

  • Repository layout, packaging, relift console script
  • 16-entry attack catalogue with risk tiers and MITRE mapping
  • list, providers, plan, scan, report, init-scope commands
  • Scope file with allowlist / denylist / CIDR / expiry / risk ceiling
  • Dry-run default; --execute required to send packets
  • Stdlib modules: DNS, TCP connect scan, HTTP fingerprint, headers, methods, robots/sitemap, TLS certificate, TLS protocol support
  • AI abstraction with Ollama, LM Studio / llama.cpp, OpenAI, Anthropic, OpenRouter, Groq and an offline rule engine
  • JSONL audit log + per-run JSON artefact
  • Exit criteria: relift list renders; scan --execute runs at least six modules against a local service; --no-ai produces the same findings as --ai none.

M2 — Depth (shipped in 0.2.0)

  • Database auditing: PostgreSQL, MySQL/MariaDB, Redis, MongoDB, MSSQL, Elasticsearch
  • Credential verification against a live server, one attempt per credential, consent-gated
  • Host audit from the local machine (users, sshd config, world-writable paths, SUID)
  • SSH transport pre-auth enumeration (banner, KEXINIT, deprecated algorithms)
  • Single-request SQL injection check (error-based + boolean), fixed payload set
  • Findings correlation: chains.py merges findings into multi-stage attack threads
  • Config file (relift.config.json) with profiles and per-engagement defaults
  • Desktop application (Windows/Linux): loopback HTTP API + token-authenticated UI
  • Severity scoring with CVSS-style vector strings from the module, not the AI
  • HTML report, single self-contained file
  • relift diff runA runB — what changed since last sweep
  • Exit criteria (remaining): two consecutive runs produce a readable diff; HTML opens offline.

M3 — Breadth

  • Subdomain enumeration from a passive source (no brute-force) as a module
  • SMB / LDAP / SSH handshake fingerprinting for the internal network
  • Discovery mode: enumerate live hosts on an in-scope CIDR before scanning
  • Parallel execution with a global request budget
  • Exit criteria: internal /24 sweep completes under the configured request budget without tripping any scope rule.

M4 — Analyst quality

  • Provider-side structured output (JSON mode) for triage, with schema validation
  • Local model recommendation table (VRAM vs. quality) in 04-AI-PROVIDERS.md
  • Prompt-injection defence: banner/title text is data, never instructions
  • Exit criteria: a hostile banner containing instructions cannot make the analyst emit a non-catalogue module or a shell string — covered by a test.

M5 — Open-source release (decision point)

  • External review of the authorisation gate
  • Relicense (Apache-2.0 or AGPL-3.0 — decision deferred to Jay)
  • CONTRIBUTING with the "no tier-D modules" rule stated up front
  • Exit criteria: a stranger can run the quickstart in ten minutes and cannot accidentally scan a host outside a scope file.

7. Risks

Risk Impact Mitigation
Tool used against a non-owned host Legal exposure, harm to others Mandatory scope file, deny-by-default, --execute flag, prominent docs
Scope file authored carelessly (0.0.0.0/0) Accidental mass scanning Range-size guard: refuse CIDRs broader than /22 without --i-know
AI hallucinates a technique or a shell command Unsafe or misleading action Catalogue-ID whitelist, no command generation, output validated before use
Local model produces confident nonsense Bad report Findings come from modules; AI only rephrases, and a --no-ai run is always possible
External tool version drift Wrapper breaks Fixed flag templates, fail soft, record the missing binary as a skipped execution
Own tool flagged by IDS/SOC on company network Noise, false alarm Tier classification + rate limits + "announce your own sweeps" note in WS-6 docs

8. Success metrics (first 90 days)

  • Zero runs against hosts not in a scope file (audit log is the source of truth).
  • list → scan → report on a fresh lab VM in under 5 minutes from clone.
  • Every finding traceable to a catalogue ID and an execution record.
  • Full offline operation verified: airplane mode, --no-ai, nothing breaks.

9. Working agreements (single-developer, but future-proof)

  • A technique is not "in" the tool until it has a catalogue entry with a risk tier.
  • Any module that mutates target state is rejected at review, not "temporarily disabled".
  • Docs describe behaviour that exists. Design-only entries say so in the same table.
  • Credits stay explicit: Jay / Pixelcity in __init__, pyproject.toml, README, CLI banner.

10. Immediate next actions

  1. Run relift list and prune anything that does not earn its risk tier.
  2. Spin a deliberately vulnerable container in the lab and run M1's exit criteria.
  3. Decide the module contract for M2 correlation before adding more modules.
  4. Pick the local model the analyst will actually use day to day.
Architecture →