efterlev 0.1.155

Repo-native, agent-first compliance scanner for FedRAMP and DoD Impact Levels

Efterlev

Open source. Runs locally. Compliance evidence lives in your repo alongside the code it describes.

Efterlev reads your AWS infrastructure-as-code — Terraform .tf and terraform show -json plan output, plus AWS CloudFormation YAML/JSON templates (graduated to default-on at v0.1.99) — classifies it against the 60 thematic Key Security Indicators of FedRAMP 20x, drafts FRMR-compatible attestations grounded in cited evidence, and proposes code-level remediations. No procurement cycle. No vendor account. Apache 2.0.

A 100-person SaaS company gets told by its biggest prospect: "we'll buy, but only if you're FedRAMP Moderate." Consulting engagements start at $250K; SaaS GRC platforms cover SOC 2 beautifully and treat FedRAMP as a footnote. Efterlev is the tool a single engineer can install on a Tuesday and bring concrete results to Wednesday's standup.

What you get

The Gap Agent classifies all 60 KSIs against your IaC and writes a color-coded HTML report + a reviewer-ready POA&M markdown:

The POA&M markdown is a real artifact 3PAOs can read directly:

# POA&M — fedramp-20x-moderate

**DRAFT — requires human review.** Severity is a starting-point heuristic
(not_implemented → HIGH, partial → MEDIUM); reviewer must confirm.

- **Baseline:** fedramp-20x-moderate
- **FRMR version:** 0.9.43-beta
- **Open items:** 32

| POA&M ID                  | KSI                                       | Status            | Severity |
|---|---|---|---|
| `POAM-KSI-CMT-RMV-000`    | KSI-CMT-RMV — Redeploying vs Modifying    | `not_implemented` | `HIGH`   |
| `POAM-KSI-CNA-RNT-004`    | KSI-CNA-RNT — Restricting Network Traffic | `not_implemented` | `HIGH`   |
| `POAM-KSI-IAM-MFA-007`    | KSI-IAM-MFA — Phishing-Resistant MFA      | `not_implemented` | `HIGH`   |
| `POAM-KSI-SVC-SNT-014`    | KSI-SVC-SNT — Securing Network Traffic    | `not_implemented` | `HIGH`   |

Sample from a real csp-starter-cfn run; full POA&M has 32 open items. See evals/PHASE_2_LITE_CFN_VALIDATION.md for the maintainer-validation numbers (23/23 = 100% precision + 100% recall on this fixture).

pipx install efterlev
efterlev quickstart                        # one-command demo against a bundled fixture
                                           # (~3 min on Sonnet, ~$0.30; runs init+scan
                                           #  only without ANTHROPIC_API_KEY set)

Or against your own repo:

pipx install efterlev
cd path/to/your-repo                       # repo root, NOT a Terraform subdir
efterlev init --target . --force
export ANTHROPIC_API_KEY=sk-ant-...
efterlev report run                        # init → scan → gap → document → poam

Pronounced "EF-ter-lev." From Swedish efterlevnad (compliance).

Or have an AI assistant do it for you

Open Claude Code, Cursor, Codex, Kiro, or any AI assistant with shell access, and paste the canonical prompt at docs/ai-quickstart-prompt.md. The assistant will confirm the repo root, ask which backend you want, install Efterlev, run the full pipeline, and brief you on the top 3 KSIs to focus on. Cost: ~$0.30–1 on Sonnet 4.6 (the recommended default), ~5–10 minutes wall.

---

Why this exists

A 100-person SaaS company just got told by its biggest prospect: "we'll buy, but only if you're FedRAMP Moderate."

The team googles it. Consulting engagements start at $250K. SaaS compliance platforms cover SOC 2 beautifully and treat FedRAMP as a footnote. Enterprise GRC tooling is priced for the wrong scale. A NIST document family runs to thousands of pages.

What they actually need is something that reads their infrastructure-as-code — whatever flavor they use — and tells them, in their own language, what's wrong and how to fix it. Something a single engineer can install on a Tuesday and show results at Wednesday's standup. Output concrete enough that their 3PAO can use it; honest enough that the 3PAO won't throw it out.

Efterlev is that tool.

It targets FedRAMP 20x — the new authorization track that replaces narrative-heavy System Security Plans with measurable outcomes called Key Security Indicators. KSIs are concrete things ("encrypt network traffic," "enforce phishing-resistant MFA") that can be assessed against actual evidence rather than long descriptions of intent. Most new SaaS authorizations starting in 2026 will target this track. Efterlev's primary internal abstraction is the KSI; FRMR (the machine-readable format FedRAMP 20x is standardizing on) is the primary output.

Built for 20x. Not retrofitted from Rev 5.

Most compliance tooling was architected for Rev 5 — narrative SSPs, manually-assembled evidence packages, GRC exports compiled around quarterly assessment windows — then bolted onto KSI workflows after FedRAMP announced 20x. That retrofit works for documentation transformation. It breaks down for persistent validation, where assessors evaluate the validation process itself — pipelines, code, automation — rather than compiled artifact packages.

Efterlev was architected for the KSI/FRMR target from day one. KSI is the primary internal abstraction. FRMR is the primary output. The detectors are deterministic and replayable; the agent prompts are plain .md files an assessor can read; the provenance chain walks back from any claim to the exact file and line that produced it. Not a Rev 5 tool wearing a 20x hat.

Compliance falls out of your security program, not alongside it.

GRC has historically been a parallel workstream — something you run next to your security program to produce compliance artifacts. Efterlev inverts that. Security signals are structured, reusable, and continuously produced; the FRMR package is the byproduct, not the goal. When you ship a Terraform change, the attestation updates. When CI runs on a PR, the KSI verdicts move. The compliance output follows the evidence automatically — continuous proof replaces periodic documentation.

---

What it does

• Scans your AWS IaC — Terraform .tf files and terraform show -json plan output, plus AWS CloudFormation .yaml/.yml/.json templates (default-on at v0.1.99; 60/60 detectors reachable from CFN, maintainer-validation 44/44 = 100% precision + 100% recall across 2 fixtures) — for evidence of 60 thematic KSIs, backed by underlying NIST 800-53 Rev 5 controls. AWS CDK Python source-mode (introduced v0.1.126; 27 supported constructs at graduation v0.1.132) emits Evidence with .py file:line citations for construct presence and inventory — composes with the existing synth-mode (cdk synth → CFN → scan) for property-level depth.
• Classifies each KSI as implemented, partial, not_implemented, not_applicable, or evidence_layer_inapplicable (the honest answer for procedural KSIs no scanner can see)
• Drafts FRMR-compatible attestation JSON grounded in cited evidence — every assertion cites its source file (and HCL line numbers when scanning .tf directly; plan-JSON mode resolves modules at the cost of file-level-only citations)
• Proposes code-level remediation diffs you can review, edit, or apply
• Generates a reviewer-ready POA&M markdown for every open KSI, with out-of-boundary scope filtering
• Traces every claim back to the file (and HCL line range, in .tf mode) that produced it via efterlev provenance show <id> — accepts truncated SHA prefixes
• Watches: efterlev report run --watch re-runs the full pipeline on every save (debounced 2s)
• Captures token telemetry so you can audit per-run LLM cost without consulting CloudWatch

Everything runs locally. The only outbound network call is to your configured LLM endpoint — direct Anthropic API by default, or AWS Bedrock ([bedrock] extra) for FedRAMP-authorized GovCloud deployments. Scanner output is fully deterministic and offline.

What it doesn't do

• It does not produce an Authorization to Operate. Humans and 3PAOs do that.
• It does not certify compliance. It produces drafts that accelerate the human review cycle.
• It does not guarantee LLM-generated narratives are correct. Every claim carries requires_review: Literal[True] at the type level — not a flag, not a string.
• It does not cover SOC 2, ISO 27001, HIPAA, or GDPR. Other tools serve those well.
• It does not scan live cloud infrastructure (yet — v1.5+).
• It does not replace AWS Config / Security Hub for runtime evaluation. Efterlev is the pre-deploy IaC layer; AWS-native is the runtime evidence layer. See docs/aws-coexistence.md.

For the honest full accounting, see LIMITATIONS.md.

---

How to run it

efterlev init --target . --force               # creates .efterlev/ workspace
efterlev boundary set \                        # declare authorization scope
  --include 'infra/terraform/**' \
  --include '.github/workflows/**'
efterlev doctor                                # pre-flight check (Python, FRMR cache,
                                               #  API key shape, Bedrock creds, LLM ping)
efterlev scan                                  # raw .tf files
# OR for module-composed codebases (the dominant pattern):
terraform init && terraform plan -out plan.bin && terraform show -json plan.bin > plan.json
efterlev scan --plan plan.json                 # ~60% more evidence on real codebases

efterlev agent gap                             # KSI-by-KSI classification (Opus 4.7)
efterlev agent document                        # FRMR JSON + HTML attestations (Sonnet 4.6)
efterlev agent remediate --ksi KSI-SVC-SNT     # Terraform diff that closes the gap (Opus 4.7)
efterlev poam                                  # POA&M markdown for every open KSI
efterlev provenance show <prefix>              # walk any claim back to source (8-char prefix OK)
efterlev provenance verify                     # tamper-evidence sweep

Or just:

efterlev report run                            # full pipeline: init → scan → gap → document → poam
efterlev report run --watch                    # re-run on every file change (2s debounce)

Wire it into CI: drop-in GitHub Action at .github/workflows/pr-compliance-scan.yml posts a sticky markdown PR comment with findings + detector coverage. See docs/ci-integration.md.

Troubleshooting

Real customer repos hit a few recoverable failure modes when generating plan-JSON. The full recovery dance lives in docs/ai-quickstart-prompt.md (Step 4); the short version:

• terraform init fails on a missing/locked remote backend (very common when scanning a repo you don't operate) — skip the remote-state machinery: terraform init -backend=false && terraform plan -refresh=false -out plan.bin. If the repo has a terraform { backend "s3" {} } block that -refresh=false doesn't bypass and you see "Backend initialization required", drop straight to HCL mode (next bullet) — don't burn 5 minutes on the dance.
• terraform plan fails on missing required variables — create a throwaway .tfvars with placeholders and pass it via -var-file. If both routes fail, drop to HCL mode.
• HCL-mode fallback — efterlev scan with no --plan flag. Keeps HCL line numbers in citations (which plan-JSON loses); the trade-off is missed coverage on resources defined inside upstream modules.
• efterlev doctor is the one-stop pre-flight check — it actively pings the LLM (Anthropic API or Bedrock InvokeModel) so credential / model-access issues surface before agent runs spend money.

---

How it's built

Three layers, each with a clear job:

• Detectors — small, deterministic Python folders. One detector = one folder = one compliance pattern. No AI. The detector library is the community-contributable surface.
• Primitives — typed functions wrapping the things agents need ("scan this directory," "validate this output," "load that catalog"). MCP-exposed.
• Agents — focused reasoning loops backed by Claude. Each has its system prompt in a plain .md file you can read and audit. AI is used for the parts where reasoning matters; never for the parts where determinism does.

This split — deterministic for evidence, AI for reasoning, different model weights for different cognitive loads — is the most important design decision in the project. It's what lets us tell auditors and 3PAOs the truth: scanner findings are verifiable facts about your code; AI claims are drafts you can audit but should not blindly trust.

The intent → execution → outcome chain. A FedRAMP 20x assessor evaluating the validation process itself traces three things as a single verifiable chain: declared intent, faithful execution, verifiable outcome. Efterlev maps directly to this. Intent is declared in the KSI catalog — concrete, measurable indicators with explicit pass/fail criteria, not narrative descriptions of "what we try to do." Execution is the deterministic detector code under detectors/ (auditable Python; one folder per compliance pattern) plus the agent prompts under src/efterlev/agents/ (plain .md files a 3PAO can read once and trust until they git-diff). Outcome is the FRMR JSON with content-addressed evidence IDs that walk back to the exact file and line range via efterlev provenance show <id> — tamper-evident via efterlev provenance verify. That's the persistent-validation chain assessors will look for under PVA-TPX-UNP: a verifiable pipeline whose intent, execution, and outcome are all readable, replayable, and bound together.

Multi-detector-per-KSI: defense-in-depth at the IaC layer. A common shallow-automation failure mode is mapping a single config check to a single KSI and calling the control implemented — a 3PAO worth their accreditation catches that immediately. Efterlev aims for multiple detectors per KSI wherever the control intent spans multiple resource types or configuration dimensions. KSI-CNA-RVP (network segmentation and traffic protection) is currently evidenced by 9 distinct detectors across WAF rules, security groups, NACLs, and load balancer listener configurations — the most cross-resource-type-evidenced KSI in the library by a wide margin. A KSI is satisfied only when control intent is evidenced across the relevant dimensions, not when any one of them happens to match.

Hallucination defenses are structural, not advisory. Every AI-generated claim links to evidence records via content-addressed IDs. Prompts wrap evidence in <evidence_NONCE> XML fences with a per-run nonce; a post-generation validator rejects any output citing IDs the model didn't actually see. The provenance store rejects any claim whose derived_from cites IDs that don't resolve. The DRAFT marker is Literal[True] at the type level — there's no flag to clear it.

Secrets never leave the machine unredacted. Every LLM prompt is unconditionally scrubbed for 7 secret families (AWS keys, GCP keys, GitHub tokens, Slack tokens, Stripe keys, PEM private keys, JWTs). The scrubber has no opt-out path. Each redaction writes an audit line to .efterlev/redactions/<scan_id>.jsonl (mode 0o600); review with efterlev redaction review.

LLM calls degrade predictably. Transient errors retry with exponential backoff + full jitter (3 attempts). On primary-model exhaustion, falls back once from Opus to Sonnet before surfacing a failure. Non-retryable errors (auth, invalid request) fail immediately. Each call's token usage is captured on the resulting Claim record and written to .efterlev/receipts.log for offline cost auditing.

For deeper architectural detail, see docs/architecture.md. For the design history including reversals and tradeoffs, see DECISIONS.md.

---

Coverage at v0.1.137

• 66 detectors — 62 KSI-mapped + 4 supplementary 800-53-only (where FRMR 0.9.43-beta doesn't yet map the underlying control)
• 37 of 60 thematic KSIs covered, across 10 of 11 themes (CMT, CNA, IAM, MLA, PIY, RPL, SCR, SVC, plus partial PIY-RSD via the github workflow detector, plus partial AFR-UCM via the cryptographic-module mapping audit in v0.1.42). The Tier 2 serverless detector batches (v0.1.44–v0.1.45) widened the resource-type evidence base across the CNA family (Lambda + API Gateway dimensions) without changing the KSI count. The remaining themes (CED, INR, plus the procedural-only KSIs in detector-covered themes) are entirely procedural — covered by customer-authored Evidence Manifests rather than detector evidence.
• Detector sources: 57 Terraform (KSI-mapped) + 5 GitHub workflows + 4 supplementary
• Three agents: Gap (Opus 4.7), Documentation (Sonnet 4.6), Remediation (Opus 4.7)
• Two LLM backends: Anthropic API (default) + AWS Bedrock ([bedrock] extra, GovCloud-deployable). Bedrock Claude Haiku 4.5 maintainer-validated at 111/112 = 99.1% precision + 100% recall across all 5 labeled fixtures (v0.1.118 — see docs/benchmark-2026-05.md); same model on Anthropic API hit 111/111 = 100%. Quality-neutral switch — GovCloud customers get the same classification quality at ~$0.40 per fixture eval.
• 2069 tests passing; mypy strict + ruff check + ruff format clean across 271 source files; 40 CLI commands; full E2E pipeline smoke (real Anthropic API call against a synthetic fixture) runs as a required check on every PR

Coverage relative to FedRAMP 20x Phase 2's 70% automated-validation threshold: the threshold applies to the customer's whole authorization package, not to any single tool. Efterlev covers 37 KSIs at the IaC layer pre-deploy (post-Tier-2 serverless backlog); AWS-native services (Config, Security Hub, CloudTrail, Inspector, GuardDuty) cover roughly 14 KSIs at the runtime layer. Honest union: ~38 of 63 KSIs (~60%) — distinct layers, not double-counted. Reaching 70% takes both. See docs/aws-coexistence.md for the strategic mapping and docs/csx-mapping.md for how the outputs map to CSX-SUM / MAS / ORD.

---

Where Efterlev fits

Sits alongside AWS Config / Security Hub / CloudTrail, not in place of them:

	Efterlev	AWS-native
When	Pre-deploy, on every commit or save	Post-deploy, on a 3-day cadence
Reads	Terraform .tf + .github/workflows/*.yml + .efterlev/manifests/*.yml + AWS CloudFormation .yaml/.json (default-on at v0.1.99)	Live AWS API state, runtime events
Output	Per-KSI attestation JSON + POA&M markdown + remediation diffs	Config evaluations, Security Hub findings, CloudTrail logs
Cost	Free (Apache 2.0); ~$0.30–2 per run on the LLM endpoint you configure	AWS spend

A FedRAMP 20x customer pursuing the 70% automated threshold typically wires both, plus procedural Evidence Manifests under .efterlev/manifests/*.yml for the AFR / CED / INR themes detectors can't see.

---

Run it from another AI session (MCP)

efterlev mcp serve

Exposes every CLI verb as an MCP tool over stdio. Point Claude Code (or any MCP client) at it and drive scans, agent calls, and provenance walks from another AI session. Our own agents use the same MCP interface — that's how we know it works. If you want to build a compliance workflow Efterlev doesn't ship, write your own agent against the MCP surface; you don't need to fork the codebase.

---

Documentation

Full docs site: efterlev.com — quickstart, concepts, tutorials (CI integration, GovCloud deployment, writing detectors, customizing agent prompts), CLI reference, and comparisons against Paramify, Comp AI, Vanta/Drata, and traditional consulting.

In this repo:

• docs/ai-quickstart-prompt.md — canonical prompt for AI-assistant-driven runs
• docs/quickstart.md — five-minute path from zero to your first scan
• docs/architecture.md — three-layer architecture in depth
• docs/aws-coexistence.md — how Efterlev fits next to AWS-native services
• docs/csx-mapping.md — outputs mapped to CSX-SUM / MAS / ORD
• docs/ci-integration.md — drop-in GitHub Action for PR compliance scans
• docs/deploy-govcloud-ec2.md — running inside an AWS GovCloud boundary
• docs/icp.md — Ideal Customer Profile; the lens behind every product decision
• docs/followups.md — tracked work, deferred items, what's shipped vs pending
• docs/v0.2-eval-harness-plan.md — design sketch for the v0.2 agent-quality harness
• CHANGELOG.md — release-by-release record (156 patch versions, v0.1.155 current)
• docs/benchmark-2026-05.md — time-to-FRMR benchmark methodology (numbers populated by maintainer dispatch)
• docs/cfn-detector-parity.md + docs/cfn-detector-parity.csv — per-detector CFN coverage matrix (regenerated by scripts/build_cfn_parity_table.py)
• DECISIONS.md — append-only design-decision log

---

Contributing

We want contributors. The detector library is designed to make the common contribution — "here's a new KSI indicator I can evidence from Terraform" — a self-contained folder that doesn't touch the rest of the codebase.

CONTRIBUTING.md has the five-minute path from git clone to running tests, the hour path from idea to open PR, and the per-fix regression-test discipline every patch ships under. Community conduct: Contributor Covenant 2.1. Good first issues are labeled good first issue on GitHub. The most valuable contributions right now are new detectors covering KSIs on the roadmap.

---

Status, governance, license

Status: v0.1.155 is current — one hundred and fifty-six patch releases since v0.1.0 (2026-04-29). The CloudFormation/CDK synth-mode arc shipped progressively from v0.1.72 through v0.1.99: parser + adapter + scan integration (v0.1.72), property-mapping table batches 1-9 covering all 60 detector-referenced CFN types (v0.1.73-v0.1.96, with the v0.1.95 detector parity audit surfacing and v0.1.96 closing 8 missing-mapping gaps surfaced by that audit), labeled eval-harness fixtures csp-starter-cfn (v0.1.78) and aws-vpc-cfn (v0.1.97), and two maintainer-validation dispatches landing at 44/44 = 100% precision + 100% recall combined (csp-starter-cfn v0.1.81 = 23/23 on Bedrock Haiku 4.5; aws-vpc-cfn v0.1.98 = 21/21 on Anthropic API Haiku 4.5) — matching the v0.1.69 Terraform-side baseline (67/67 across 3 fixtures at 100/100). CFN graduated to default-on at v0.1.99; the --allow-cfn flag is deprecated for one minor release (removed in v0.2.0). The strategic bet is validated end-to-end: no property-mapping or detector bugs surfaced; classification quality is invariant under IaC syntax. Detailed per-KSI analysis: evals/PHASE_2_LITE_CFN_VALIDATION.md. Per-detector CFN coverage: docs/cfn-detector-parity.md + docs/cfn-detector-parity.csv. Per-release detail in CHANGELOG.md; verify a published artifact with bash scripts/verify-release.sh v0.1.155 (PEP 740 PyPI attestations + cosign keyless OIDC + SLSA provenance on ghcr.io/efterlev/efterlev).

Governance: Benevolent-dictator model today (@lhassa8), transitioning to a technical steering committee at 10 sustained-activity contributors. Full model in GOVERNANCE.md. Architectural decisions: DECISIONS.md. The project may eventually be donated to a neutral foundation (OpenSSF / Linux Foundation / CNCF) if contributor diversity warrants — that decision is not made and not time-boxed.

License: Apache 2.0. See LICENSE.

Security: Coordinated disclosure process in SECURITY.md. Threat model for Efterlev itself: THREAT_MODEL.md. The pre-launch security review (signed by the maintainer) is at docs/security-review-2026-04.md.

---

Credits

Efterlev was bootstrapped using Claude Code. The architecture commits to keeping Claude Code (and other MCP-capable agents) as first-class integration partners — that's what "agent-first" means here, structurally, not as marketing.

Built on compliance-trestle for OSCAL catalog loading, on the FedRAMP Machine-Readable (FRMR) catalog, and on the NIST SP 800-53 Rev 5 catalog. Those projects make this one possible.