llm-ingestion-pipeline-security/README.md

llm-ingestion-guard

A reusable, minimal, dependency-light defensive layer for LLM ingestion pipelines — the write-time siblings of query-time chatbot guardrails.

Where mature guardrails (LLM Guard, NeMo Guardrails, Rebuff, Vigil, …) sit between a user and a model at query time, this library hardens the other shape: untrusted content flowing through an LLM enrichment/summarization/extraction step into a persisted, downstream-consumed artifact (RAG corpus, knowledge base, wiki). It packages the architectural contract — sanitize → fence → tool-less quarantined transform → per-stage capability isolation → scan output before commit → fail-secure — as composable, stdlib-first, framework-agnostic code.

The gap it fills is not "no one detects injection." Existing OSS tools are either single-stage detectors — they emit a risk verdict and leave quarantine, capability isolation, scan-before-persist, and fail-secure disposition to the integrator — or runtime content-hardening that guards material as it enters the model. This library packages the full write-time ingestion contract as composable code: the part query-time tooling structurally cannot see, because a poisoned artifact committed at write time is read by a downstream agent whose guardrail never sees where it came from. (Nearest neighbours surveyed in docs/BRIEF.md §11.)

Status: v0.2, alpha. The stdlib-only core is built and tested — ten detector/contract modules and the top-level wiring, exercised by an end-to-end showcase and adversarial + false-positive corpora. The public API may still change. There are real limitations, stated plainly below; read them.

Install

pip install llm-ingestion-guard          # stdlib-only core, zero dependencies

Optional ML/judge detectors live behind extras ([ml], [judge]) and are not required — the core is deterministic and dependency-free.

Quickstart — the two bookends

The library never makes the model call itself. It gives you the two library-side halves around your own tool-less transform:

from llm_ingestion_guard import (
    prepare_input, screen_output, Disposition, PRESET_USER_UPLOAD,
)

prepared = prepare_input(untrusted_content)              # §6 1-2: sanitize + fence
enriched = your_model(prepared.fenced)                   # §6 3: tool-less — YOUR call
decision = screen_output(enriched, PRESET_USER_UPLOAD)   # §6 6-7: scan + dispose

if decision.disposition is Disposition.FAIL_SECURE:
    alert(gate_code=decision.reasons)                    # §6 8: minimal payload, no content
    raise SystemExit                                     # §6 7: halt — never persist

screen_output fails closed: if the scanner itself errors on crafted input, the disposition is FAIL_SECURE, never a silent persist. Pass transform_failed=True when your model call raised or fell back — a scan hit together with a transform failure is treated as a probable forced-fallback attack and halts regardless of trust tier.

Every primitive is also exported for pipelines that compose the checklist themselves — sanitize, scan_lexicon, scan_entropy, scan_output, scan_active_content, neutralize, the decide / guard disposition machinery, and the contract asserters assert_tool_less / assert_credential_allowlist / scoped_env. See the end-to-end showcase for a full worked pipeline.

Verify what it stops (coverage matrix)

Before wiring the guard into your pipeline, see — in one place — every vulnerability class it stops, and the ones it deliberately does not:

python -m llm_ingestion_guard.coverage    # prints the matrix; exit 0 = all as documented

Each row drives the real guard with a live payload and prints class -> OWASP anchor -> expected -> observed -> verdict. The manifest (coverage.py) is the single source of truth for tests/test_coverage_matrix.py, which asserts total recall across every defended class, that every documented gap still holds (a closed gap fails the test, forcing a doc update), and — the honesty guard — that every lexicon pattern has a case so the matrix cannot fall behind the lexicon. The full LLM02 secret-egress set and the container-layer front-end (CSV formula-injection, zip-slip/bomb, symlink) are asserted there too.

The reusable contract (adopt-this checklist)

The actual product is this checklist, encoded as code you wire in order:

1. Sanitize before fence. Strip carrier classes (zero-width, BIDI, Unicode-tag, HTML comment, data:) from untrusted input first.
2. Fence untrusted input. Spotlight-mark it in a randomized per-call delimiter; strip attacker fence markers from the payload.
3. Tool-less transform. Call the model with zero tools. A successful injection then has nothing to act with.
4. Per-stage capability isolation. The enrichment stage holds only the model key; the publish stage holds only the publish credential; no stage holds both.
5. Treat output as data. Parse to a frozen schema; reject on structural violation. The output never reaches a shell, git, or a filesystem path.
6. Scan output before persist. Run the lexicon + entropy + active-content scan over the emitted text. Verbatim-carried payloads, model-emitted instructions, and zero-click exfil carriers (EchoLeak-class markdown images/links, raw active HTML, data: URIs) are caught here; neutralize additionally defangs them, opt-in.
7. Fail-secure on compound signals. Injection hit + transform failure = halt
   • alert, never a silent verbatim commit.
8. Minimal alert payloads. Alert with a gate code + run ID, never content.

Steps 1-2 are prepare_input; steps 6-7 are screen_output; steps 3-5 are yours; the contract asserters harden step 3-4.

Honest limitations (shipped as a control)

Conceding these plainly is itself a control — it prevents the false assurance that a green scan means safe content:

• Structural unsolvability at the text layer. Pattern/lexicon detection is bypassable in isolation; character-injection and novel phrasings evade it. The contract (tool-less transform, capability isolation, fail-secure) is what carries the security — the lexicon is defense-in-depth, not a wall.
• A lone HIGH finding in trusted prose disposes to WARN, not quarantine. Under PRESET_TRUSTED_SOURCE, §4.7 trust-scaling downgrades a single HIGH to WARN (a reputable single-author source is expected to carry security vocabulary), and one HIGH is not "compound" — compound escalation needs ≥2 findings at MEDIUM+. So a HIGH-severity injection reproduced verbatim in output under a trusted policy persists with only a WARN. This is the §4.7 design, not a miss: if your "trusted" sources can carry attacker-influenced text, run them as untrusted (or add a quarantine floor) instead.
• The upload preset's quarantine floor is currently vacuous. PRESET_USER_UPLOAD sets quarantine_default ("any finding → at least QUARANTINE_REVIEW"), but every detector emits CRITICAL/HIGH/MEDIUM only (no LOW/INFO), and under untrusted trust a MEDIUM already escalates to QUARANTINE_REVIEW. The floor therefore never changes an outcome in today's severity set — it is defensive headroom for a future LOW/INFO finding (e.g. a grounding "unchecked" marker), not an active control. Documented so the preset's guarantee is not over-read.
• Semantic / factual poisoning is invisible to lexicon + entropy: a factually false claim in clean prose carries no suspicious token. The grounding module ships only a SourceGroundingCheck seam — the deterministic core does not judge semantics; a [judge] implementation must be plugged in.
• Adversarial-ML evasion can survive normalization; tokenizer mismatch between scanner and model leaves gaps.
• Latent / dormant memory poisoning is not judgeable at write time.
• Dormant / broken-link injection in a linked corpus (e.g. an OKF bundle): a link to a not-yet-existing target passes a per-concept write-time scan clean — the payload is planted later, when that target is written. A scanner that sees one document at a time cannot catch it; it needs cross-write graph re-scan (tracked for the OKF adapter).
• OKF reserved files (index.md / log.md). In a received bundle these are legitimate structure (directory listing, update log), so mode-b import_bundle scans their body and frontmatter — an injection planted in a directory listing is caught — rather than path-rejecting the whole conformant bundle. The upload front-end keeps the opposite rule: an individual upload landing on a reserved basename is a shadow of the listing and is refused (allow_reserved=False).
• A document that describes attacks is a false positive. Content whose legitimate purpose is to document prompt-injection payloads (security notes, this project's own corpus) trips carrier-strip / fail-secure. At the text layer there is no way to distinguish "about an attack" from "carrying an attack"; such content needs a deliberate, explicitly-marked escaped path, never a silent allow.
• Bilingual text trips the Cyrillic/Latin homoglyph rule. homoglyph:cyrillic-latin-mix (MEDIUM) flags any Latin letter adjacent to a Cyrillic look-alike, so genuine bilingual prose (e.g. Russian, or mixed Norwegian/Cyrillic) trips MEDIUM → under untrusted → QUARANTINE_REVIEW — a real false positive for an inbox that expects multilingual content. A calibration fix is pending (require ≥N mixed pairs, or only flag when a folded variant also hits another pattern); until then, multilingual corpora over-quarantine on this rule.
• Insider in-place edits by a trusted author are out of the untrusted-content threat model.
• Text-only. The core is text -> findings: it parses no files (no pypdf/python-docx/archive deps). Extract text first, then scan it with the high-untrust upload provenance. OCR-embedded instructions and multimodal stego in images/PDFs are out of scope beyond the sanitizer's character-layer stripping.
• Uploaded files: only the extracted text is scanned. The two-stage OKF inbox showcase (tests/test_okf_inbox_uploads.py + tests/inbox_frontend.py, a dev-scoped demonstration whose python-docx/python-pptx parsers live in the [dev] extra, never core dependencies) reads .txt/.md/.csv/.docx/ .pptx/.xlsx, folders and .zip, materializes them into an OKF bundle, then guards it. What survives text extraction is out of scope: VBA/macros (.docm/.xlsm/.pptm), OLE / embedded objects, image-embedded instructions needing OCR, font/render steganography, and encrypted / password-protected files — the binary layer needs a separate scanner. The front-end owns the container threats it can see (zip-slip → path gate, zip-bomb → size cap, symlink refusal, CSV/XLSX formula-lead cells). .pdf is a deliberate concession, not a TODO: a top-level .pdf drop is refused as an unsupported format rather than half-scanned, because a PDF parser (plus reportlab solely to author white-on-white test fixtures) is disproportionate for a dev-scoped showcase, and the OCR / font-render stego carriers a PDF would smuggle are already out of scope above. The one remaining known gap is the numeric -/+ CSV false positive (an XLSX numeric cell is typed, so it does not trip the gate).
• Lexicon findings are deduplicated by pattern id — count=1 and the first offset are reported, so the same class matched across several channels/variants collapses to one finding at its first location. This keeps reports readable, but a caller that counts occurrences or needs every offset of a repeated pattern sees only the first: a deliberate readability tradeoff, not full positional coverage.
• Secret egress: base64-wrapped is caught, hex-wrapped is not. The output gate decodes base64 blobs and re-scans the plaintext through the credential/egress set, so a base64-wrapped secret surfaces as decoded:egress:* rather than vanishing. A hex-wrapped secret does not: entropy exposes decoded plaintext for base64 only, so hex (and other encodings, or nested wraps) is a deliberate boundary, not a silent miss — decode the transport layer first if you need it scanned.

Out-of-scope (documented boundary)

Embedding/vector-layer defenses (OWASP LLM08, downstream of persist); multimodal steganography; query-time / runtime guardrails; semantic factuality verification.

Design & threat model

• Design brief — what this repo contains and why.
• Build plan — module build order and the reuse map.

The contract is extracted from a working reference implementation (the claude-code-llm-wiki Stage B enrichment pipeline). Threat-model anchors: OWASP LLM Top-10 2025 (LLM01/02/04/05/06 strongest, LLM08 boundary, LLM09/10), PoisonedRAG, guardrail-evasion (arXiv 2504.11168), EchoLeak (CVE-2025-32711).

License

MIT — see LICENSE.