The pptx payload hides where an audience watching the slides does not look. The
extractor surfaces all three regions into the concept text so the stage-2 scan
catches the injection:
- speaker notes (slide.notes_slide.notes_text_frame.text);
- off-slide (off-canvas) text boxes (shape.text_frame.text, position-agnostic);
- image/shape alt-text (cNvPr@descr, read off the XML — python-pptx 1.0.2 has no
stable public accessor across shape types).
Detach-proof: same visible slide without notes ADMITs. python-pptx imported
lazily (dev/showcase-scoped, not a core dep). Tests 305 -> 310.
The docx payload hides where a human reviewing the file in Word does not look.
The extractor surfaces all three regions into the concept text so the stage-2
scan catches the injection:
- hidden/vanish runs (w:vanish) — still runs, so paragraph.text includes them;
- review comments (doc.comments[].text);
- core metadata properties (subject/keywords/comments/title/category/author).
Detach-proof: the same visible body WITHOUT the hidden run ADMITs, so it is the
extractor surfacing the hidden region that caught it, not merely 'a docx'.
python-docx is imported lazily (dev/showcase-scoped, not a core dep). Tests
300 -> 305.
- .csv: cells leading with =/+/-/@ (leading whitespace stripped first) are the
CSV-injection/DDE vector the guard cannot recognize, so the front-end refuses
them; the raw cell text is still materialized so a prompt-injection phrase in a
cell is caught by the stage-2 scan (T1). Detach-proof: plain-cell version of
the same file ADMITs. Numeric -/+ leads are the accepted FP (honest-limits).
- folder: walked member-by-member with relative paths preserved, so a reserved
basename member (index.md) lands on the guard's T4 gate; symlinks refused.
Refactor: per-file dispatch shared by top-level drops and folder walk (strict
raises on unsupported top-level suffix, folder skips). Tests 293 -> 300.
The front-end reads zip entries in memory (never extracts to disk), so it owns
the container caps while a traversal entry maps onto the guard's path gate:
- zip-slip: a '../../evil.md' entry materializes onto a traversal concept path
(preserved verbatim, not normalized) -> stage-2 T4 -> FAIL_SECURE -> REJECT.
- zip-bomb: per-entry + per-archive uncompressed-size caps (OWASP LLM10) refuse
an oversize entry before its bytes are read; a bounded read defends a lying
header. Detach-proof: a generous cap admits the same archive, so the cap is
load-bearing.
- symlink entry: refused at the front-end (no legitimate concept meaning).
Caps are kwargs on extract_inbox/receive (small in tests, generous by default).
Tests 288 -> 293.
Realistic-upload showcase (PLAN §247), first slice. A stage-1 front-end reads
dropped files and materializes them into an OKF bundle {concept_path: text} +
provenance; receive() wires extract -> import_bundle -> verdict. This slice
covers .txt/.md (stdlib only); .zip/.csv/folder/.docx/.pptx follow.
- Front-end lives in tests/ (showcase/dev-scoped), core stays stdlib-only:
dependencies=[] untouched; python-docx/python-pptx added to the [dev] extra
(used from stage 2d/2e), never a public [extract] extra (not v1 per PLAN).
- .txt injection -> guard T1 -> REJECT; clean .txt -> ADMIT; a dropped .md keeps
its frontmatter so a dangerous value -> T2 REJECT; an index.* upload
materializes onto the reserved uploads/index.md -> T4 REJECT.
- Detach-proof: neuter extraction to an empty bundle -> the poisoned upload
ADMITs, proving the verdict depends on extraction carrying the payload.
Tests 282 -> 288.
The OKF analogue of tests/test_showcase.py: one received external bundle plants
one attack per OKF surface (T1 body + frontmatter-description, T3 non-https
resource, T4 path-traversal + reserved-name, T2 dangerous frontmatter value,
T5a dangerous-scheme link, §7.2 dangling link, homoglyph-obfuscated body) run
through the public okf surface as an upload-inbox consumer would compose it.
Aggregate fails secure -> REJECT; a clean bundle admits (WARN); the log marks
rejected concepts. Detach-proof: an always-admit gate catches none of the
surfaces, so the assertions have teeth. Stream-2 centrepiece (PLAN §212);
Stage-1 realistic upload extraction (§247) is the next increment.
Tests 277 -> 282.
Ground-truth pass over the named flagship consumer (portfolio-optimiser's
'OKF-upload-inbox') found it does not exist as a seam: both optimiser siblings
are frozen at release, carry their own OKF layer, receive no external bundles,
and take no dependency on this guard. Defer consumer integration (stream 2)
until the guard is mature; mature it here first, Node-port-friendly. Next
concrete build is the in-repo OKF inbox showcase (mode-b import_bundle as a
receive/quarantine gate), spec'd under 'The OKF inbox showcase'.
Google's Open Knowledge Format (OKF v0.1) is the LLM-wiki pattern this
library guards. Capture a security brief for OKF ingestion as tracked
future work -- NOT shipped scope. The v0.1 core is format-agnostic text;
OKF surfaces beyond the body (YAML frontmatter, resource URLs, cross-link
graph, file path / reserved names, format-level authenticity) are uncovered.
- Coverage claims verified against the code at 5397ba1 (brief section 9): no
YAML parse in src, neutralize defangs but has no reject-gate, no path/graph
logic, disposition machinery exists.
- OKF confirmed real at the format level (Google Cloud, 2026-06-12; spec
GoogleCloudPlatform/knowledge-catalog).
- Design move: an OKF adapter ON TOP of the format-agnostic core; the core
stays text->findings.
- Section 8 lists 8 v0.2 tasks (T1-T8); T8 (surface two OKF residual risks in
README honest-limits) can land independently.
Pre-release hardening from an independent adversarial review; each fixed
test-first (failing test -> fix -> green). 214 tests pass.
- entropy (M1): decode-and-rescan now runs BEFORE false-positive suppression,
so an SRI/media-prefixed injection blob is still decoded and lexicon-rescanned.
Suppression gates only the entropy finding, never the decode.
- output/disposition (M3): the invisible-carrier invariant now holds on the
persist gate. scan_output flags zero-width/BIDI presence and disposition
treats those + lexicon:unicode-tags-present as any-tier carriers, so a carrier
in model output fails secure even under a trusted policy.
- contract (M2): assert_credential_allowlist catches a bare <PROVIDER>_KEY
(e.g. STRIPE_KEY) that the old regex silently missed (fail-open). Deliberately
broad: also flags PARTITION_KEY/SORT_KEY as loud, allowlistable FPs -- fail-loud
beats fail-silent for an isolation control.
- disposition (m6): guard runs decide inside its guarded block -> total
fail-closed even on a malformed report.
- output (m4): egress placeholder suppression anchors word markers (example,
todo, ...) to a word boundary, closing a fail-open where a real secret merely
containing such a word was suppressed.
Docs: CHANGELOG Security subsection; README honest-limit for lexicon dedup (m5,
documented tradeoff, not fixed).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HyRCQMocjZ6SmSQ6JidJ2k
Mirror the badge convention from the ms-ai-architect plugin README — shields.io
STATIC badges (img.shields.io/badge/), which work on Forgejo where dynamic
build-status endpoints do not. Type-adapted for a library rather than a plugin:
the Platform slot becomes python-3.10+, the content-count slot becomes
tests-201-passing, plus an honest status-alpha badge.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HyRCQMocjZ6SmSQ6JidJ2k
Module 11's final deliverable (BRIEF §9, PLAN §132), built last. Exercises the
full public toolkit as a consumer would compose it — the _ingest helper doubles
as the README worked example.
- test_showcase: ONE poisoned artifact planting a vulnerability per detection
channel — 5 invisible carriers (zero-width/BIDI/unicode-tag/html-comment/
data-uri), injection across raw + homoglyph + rot13 + whole-string-base64
(decode-and-rescan) channels with DISTINCT lexicon ids so dedup won't merge
them, plus a credential the mock transform leaks into its output (LLM02). All
10 planted labels caught; disposition FAIL_SECURE; egress evidence never
carries the secret value; a clean document + clean transform WARNs (no FP).
- test_corpus: adversarial recall == 100% over the planted classes (a silent
drop is a regression), and a false-positive corpus of attack-resembling legit
content (injection vocabulary, secret-shaped placeholders/varrefs, high-entropy
checksums) that disposes WARN under a trusted source — most producing NO
finding (the suppression rules hold). One MEDIUM homoglyph case proves hard-fail
is an explicit opt-in: same finding WARNs trusted, QUARANTINEs under upload.
Secret fixtures assembled from fragments (gitleaks-safe); every payload verified
against the real detectors before assertion. 12 new tests; 201 green total.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HyRCQMocjZ6SmSQ6JidJ2k
Module 11 (final) — the top-level wiring. The library never makes the model call
(no SDK imported by the core), so the public surface is the toolkit plus two
library-side bookends around the caller's tool-less transform (Form 1, chosen
with the operator over an export-only toolkit and a full orchestrator — the
bookends fit existing pipelines with least friction, encode the two halves the
library can stand for, and impose no control flow):
- prepare_input(text, source=INPUT) -> PreparedInput(fenced, nonce, report):
§6 steps 1-2, sanitize THEN fence (carrier can never smuggle a forged
delimiter). Merged report carries both steps' findings; renders no disposition.
- screen_output(text, policy, *, provenance, transform_failed) -> DispositionResult:
§6 steps 6-7, scan_output under guard() so a scanner error fails CLOSED
(FAIL_SECURE, never a silent persist). transform_failed routes the compound
forced-fallback rule.
- __all__ exports the full framework-agnostic surface: detectors, result types,
disposition machinery + presets, contract asserters, the grounding seam.
Docs: README refreshed from the stale "brief / pre-implementation" line to a v0.1
alpha status with a Form-1 quickstart, the §6 adopt-this checklist, and an honest
-limitations section (structural unsolvability at the text layer; semantic
poisoning invisible to lexicon+entropy; text-only, no multimodal). CHANGELOG
seeded; CLAUDE.md remote/status lines corrected (remote IS set, no longer
brief-stage).
10 wiring tests (public surface, prepare_input compose, screen_output fail-closed
+ compound). 189 green (showcase + corpora follow).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HyRCQMocjZ6SmSQ6JidJ2k
Module 10 of the build order — interface only in the stdlib core (PLAN §67/§94,
BRIEF §7/§10). lexicon and entropy see surface signals; they are structurally
blind to semantic poisoning — a factually false or subtly biased claim, in clean
prose, carrying no suspicious token. Catching that needs grounding against a
source of truth (embedding classifier, retrieval check, LLM judge) — a model
call, and models never live in this core.
So this module ships exactly two things and no detector logic:
- SourceGroundingCheck: the text -> Report protocol a [judge] implementation
must satisfy to plug in. runtime_checkable for a coarse callable-vs-not
isinstance gate; a conforming impl may be a plain function or a stateful
callable holding a retriever/client — the grounding source of truth is the
impl's concern, never a seam parameter. Its findings flow through disposition
like any other detector's, with zero grounding-specific plumbing.
- no_grounding_check (bound as DEFAULT_GROUNDING_CHECK): the pass-through
default returning an empty Report. The honest scope note made structural —
semantic poisoning is not solved at the text layer, and the core does not
pretend to.
Design choice, tested: the default returns NO finding (not an INFO "unchecked"
marker). An INFO finding would flip every artifact to found and trip the
quarantine_default floor in disposition, quarantining every upload for a check
that never ran. The seam stays silent; the honesty lives in the contract.
8 new tests (pass-through empties, silence on semantic poison, protocol
conformance incl. a stateful judge stub, and a composition test proving a
plugged-in check's HIGH finding fails secure through disposition unchanged);
169 green total.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HyRCQMocjZ6SmSQ6JidJ2k
Module 9 of the build order — the differentiator. Where the other modules
detect and report, these are pure functions that ENFORCE an invariant and
raise: the write-time analogue of runtime least-privilege. They harden the
call a pipeline makes around its quarantined transform; the library makes no
model call itself.
Three asserts, matching the reusable-contract checklist (BRIEF §6 steps 3-4):
- assert_tool_less(request): the model request carries no tool surface —
tools/functions/tool_choice/function_call/mcp_servers, populated. Covers
both Anthropic and OpenAI request shapes; empty/None is genuinely tool-less.
- assert_credential_allowlist(env, allowed): the process env holds no
credential beyond this stage's allowlist (subset = least-privilege; the
enrichment stage sees only the model key, never the publish credential).
- scoped_env(env, allowed): the isolation primitive — strips off-allowlist
credentials, keeps PATH etc.; the assert then passes by construction.
Generalized from claude-code-llm-wiki tools/wiki_ingest/enrich.py
assert_quarantine (a pipeline-specific gate) into framework-agnostic pieces.
Two deliberate safety choices inherited from the reference: credential
detection is name-based (env VALUES are never read — value-scanning is the
output module's job); and a raised ContractViolation names only the offending
KEY, never a value, so it is safe to route to an alert channel (minimal-alert,
BRIEF §6 step 8). CREDENTIAL_NAME_RE ported verbatim from the proven reference.
23 new tests; 161 green total.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HyRCQMocjZ6SmSQ6JidJ2k
Module 7 of the build order: the last gate before model output is persisted.
Composes scan_lexicon + scan_entropy over the emitted text, feeds each base64
blob entropy decoded back through the lexicon (decode-and-rescan → decoded:*
findings with blob offset), and adds the LLM02 secret/credential egress layer
(cloud/provider keys, PEM headers, DB conn-strings, JWT, labelled
password/secret/api-key assignments) with placeholder/varref FP-suppression.
Report-only (never mutates; neutralize is the separate opt-in mutator). Evidence
carries only a description + match length — never the secret value. Self-safe:
single input-size cap, linear egress patterns. PEM patterns use -{5} form so the
module itself never trips a secret scanner. 26 new tests; 108 green total.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HyRCQMocjZ6SmSQ6JidJ2k
Build-order step 6. Close the EchoLeak class (CVE-2025-32711): active content in
persisted model OUTPUT that a downstream renderer auto-fetches or makes clickable,
exfiltrating data zero-click. These carriers are neither injection strings nor
high-entropy, so lexicon + entropy miss them — a distinct control (OWASP LLM05,
Improper Output Handling).
Defang, don't delete. URLs in active-content position are rewritten to a
non-resolvable but auditable form (https://evil.com -> hxxps://evil[.]com;
data:/javascript: colon neutralized to [:]); raw active HTML is escaped so a
renderer shows inert literal text. Visible information survives review; only the
machine-actionable affordance dies. Dot-defang is idempotent (never [[.]]).
Six classes, each a Finding: markdown-image (HIGH, the zero-click primitive),
inline-link (MEDIUM), reference-link definition (MEDIUM, the documented
image-filter bypass), angle-bracket autolink (MEDIUM), raw active HTML (HIGH,
inherently-active tag OR event/URL attribute — benign <b>/<em> left untouched),
standalone data: URI (HIGH). Processing order prevents double-counting.
Opt-in and separate: calling neutralize() IS the opt-in to mutate; the report-only
gate stays pure (design principles 3 & 4). Byte-identical on clean output, mirror
of the sanitizer invariant. Scope conceded in the docstring: a targeted defanger,
not a full HTML sanitizer.
17 tests: byte-identity + FP guards (lone <>[], metadata:, benign HTML), image
defang + non-resolvability, secret-exfil URL, inline/reference/autolink, raw-html
escape + script neutralization, data: URI, no double-count, counts, source.
[skip-docs]: README positioning + honest-limitations remains the deliberate
build-order step-11 deliverable (steps 1-5 likewise left README frozen).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HyRCQMocjZ6SmSQ6JidJ2k
Build-order step 5. Wrap untrusted content in a quarantine fence a downstream
trusted prompt can rely on: everything between the delimiters is data, never
instructions.
Two load-bearing properties:
- Per-call cryptographic nonce (secrets.token_hex, 128-bit) in the delimiter, so
an attacker embedded in the payload cannot forge the matching closing marker to
break out — the nonce is unpredictable and fresh every call.
- Marker-strip FIRST: fabricated fence markers already in the payload are removed
(any/no nonce, case-insensitive, ReDoS-safe negated-class regex) and flagged as
fence:marker-injection (HIGH, LLM01) before wrapping — defense in depth against
a lucky guess of the static skeleton, and it surfaces the attempt.
Pure text -> (fenced_text, report, nonce); only mutation is the marker-strip.
Nonce exposed so the caller can reference the fence in the trusted prompt.
10 tests: wrap/preserve, per-call randomness, nonce length, breakout containment,
marker-strip (+ case-insensitive), prose-word FP guard, source, empty input.
[skip-docs]: README positioning + honest-limitations is a deliberate build-order
step-11 deliverable (steps 1-4 likewise left README frozen). README status line
("pre-implementation") is stale and flagged for the step-11 refresh.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K8GmKRCdsPjWYAKWsNgeQS
One realistic content sample carrying many vulnerabilities at once, run through
a mock ingestion pipeline (sanitize -> lexicon+entropy+decode-rescan -> output
gate -> disposition); assert every planted vuln is caught and disposition fails
secure. Doubles as the README worked example. Added to build-order step 11 and
the test strategy; tracked as task #11. Inspiration: llm-security/examples/*.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K8GmKRCdsPjWYAKWsNgeQS
Build order step 4 — the load-bearing port from the llm-security seed
(injection-patterns.mjs + string-utils.mjs), stdlib-only.
- injection_lexicon.json: 83 patterns (CRITICAL/HIGH/HYBRID/MEDIUM) as the
single source of truth (regex + id + severity + owasp + desc), compiled once
by a thin loader. Decoupled from the engine for a future TS port.
- scan_lexicon(text, source, max_scan_chars) -> Report: matches every pattern
against a deduped variant set (raw / normalized / homoglyph-folded / rot13),
plus unicode-tag presence signal and the cognitive-load trap.
- normalize_for_scan chain ported: unicode-tags -> bidi -> HTML-entities ->
unicode/hex/URL escapes -> whole-string base64 (reuses entropy.try_decode_base64)
-> collapse letter-spacing; plus fold_homoglyphs / rot13.
- Self-safety (OWASP LLM10): input-size cap (scan prefix + flag oversize) and
ReDoS-safe port — the two nested-.*? sub-agent patterns bounded to
(?:\S+\s+){0,N}?; verified true positives still fire.
- Non-Latin data (homoglyph map, BIDI block) built from explicit code points;
JSON non-ASCII kept as \uXXXX escapes.
24 tests; 55 green total.
[skip-docs]: README positioning + honest-limitations is a deliberate build-order
step-11 deliverable (steps 1-3 likewise left README frozen). README status line
("pre-implementation") is stale and flagged for the step-11 refresh.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K8GmKRCdsPjWYAKWsNgeQS
Build order step 3. Pure text -> findings detector ported from the
llm-security entropy-scanner seed:
- Length-calibrated Shannon-entropy classification (CRITICAL 5.4/128,
HIGH 5.1/64, MEDIUM 4.7/40).
- Shape floor: base64-like (len>100) / hex (len>64) reach at least MEDIUM
even when entropy alone does not trigger — the only path that catches hex
(16-symbol alphabet caps H at 4.0 < 4.7).
- Decode-and-rescan (must-have): base64 blobs that decode to printable text
are exposed on EntropyResult.decoded for a later lexicon rescan.
- FP suppression scoped to the text-relevant subset (base64 media data-URI
prefixes + SRI sha*- prefix); source-code-specific seed rules omitted.
Exposes ported primitives shannon_entropy / is_base64_like / is_hex_blob /
try_decode_base64. 16 new tests; full suite 31 green. Stdlib-only.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K8GmKRCdsPjWYAKWsNgeQS
Add .gitignore (Python stack + obligatory lines; /STATE.md kept local-only
until a private remote exists), CHANGELOG.md, and the Communication patterns
section in CLAUDE.md. STATE.md is gitignored — must never reach the public
open/ mirror.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K8GmKRCdsPjWYAKWsNgeQS
Based on the working, tested Layer B ingestion-gate in ms-ai-architect (2026-07-04):
- A: correct consumer 2 — it fetches authored Learn docs + code samples via the
microsoft-learn MCP, NOT the open Q&A forum / MSDN / Stack Overflow (that claim
was unverified/overstated); low-trust surface is intra-document. Repo/name
reconciled: the consumer is ms-ai-architect, no separate "MS AI Security plugin".
- B: §4.7 — trust tiers WITHIN a document (code sample / localized string vs
authored prose), not only across sources; carriers + critical block in any tier.
- C: §12 — consume the lexicon as pure imported functions, not the llm-security
scan CLI (which under-covers Markdown prose + base64-in-code-block, verified).
- D: §4.6 — fail-secure extends to scanner-unavailable: un-scannable ⇒ BLOCK.
Adds ms-ai-architect as the second reference implementation (output side).
Second target consumer named: MS AI Security plugin, an ingestion
pipeline over Microsoft Learn content that includes user-generated Q&A
(learn.microsoft.com/answers, plus ingested MSDN/Stack Overflow). This
is the high-untrust case where the contract is load-bearing, not
hygiene. New design principle 4.7: disposition scales with source trust
(pinned changelog = WARN; open UGC = quarantine/hard-fail on high
severity). Day-1 rationale: architectural controls are cheap to design
in, expensive to retrofit (consumer 1 is proving that at its A13).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HPAmFyEVWbwvmSNVdXTu4d
