First consumer swap of step 4. ZERO_WIDTH_CHARS (5), the Unicode Tag range,
BIDI_CHARS (9) and HOMOGLYPH_MAP (28) stop being hardcoded constants in
unicode-scanner.mjs and string-utils.mjs and are built from the vendored
commons artifact codepoints/carriers.json by the new lib/codepoints.mjs.
Started here rather than at injection-patterns, which the plan ordered first:
that table is the one table that cannot be loaded verbatim (the
hybrid-xss:script-tag divergence is directional, and loading the lexicon as-is
would reverse the 90f576f recall fix). The codepoint tables were measured
byte-equal to the source constants BEFORE the swap - same members, same
values, same insertion order on HOMOGLYPH_MAP - so they load verbatim.
Proof the swap is content-preserving: the golden dump differs in exactly one
record, the sha256 of string-utils.mjs, which changes by construction when a
table leaves the file. All 83 regex records and the
table:string-utils:HOMOGLYPH_MAP digest are byte-identical, and
reference-run.json is unchanged at 61/61. patterns.json is re-blessed for the
file digest alone.
The gate is proven red-capable against the SUBJECT, both directions:
- dropping U+00AD from the vendored zero_width table fails the new
codepoints gate by name, twice;
- altering one homoglyph value reddens the golden table digest AND a
behavioural homoglyph test.
That second direction is a property the swap creates rather than preserves:
the golden gate now transitively pins the vendored commons data, where before
it pinned a source literal and a commons mutation was invisible to it.
NOT ported: commons carries cyrillic_confusables (13), and unicode-scanner.mjs
declares a set by that name - but nothing reads it. The homoglyph-mixing
detector tests isCyrillic(cp), the whole U+0400-U+04FF block. Loading it would
move dead data into the load path, so the dead const stays where it is and is
recorded instead. The recorded v8.x-B i/x drift between that set and the
lexicon class is therefore latent, not live. commons' private_use table has no
constant behind it here at all.
Graceful-empty is kept deliberately: codepoints.mjs is on string-utils'
import path and hooks import string-utils in fresh per-tool-call processes, so
a module-load throw would break the tool call rather than degrade the scan.
The loud half is the test, which asserts exact per-table counts through the
real default commons root - the same shape as the lexicon load-assertion.
Drive-by, unavoidable: the deleted JSDoc carried the "~25 entries" claim for a
28-entry table (v8.x-C). It needed a re-bless of the same file digest this
swap already forces, so it closes here at no extra cost.
Suite 2158 -> 2164, all green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X7XEEFrAJsREqa9N4tpfm8
Vendored commons v0.1.0 carries a cross-runtime conformance corpus. Until now
it was measured by a throwaway script, which makes 83/83 a claim rather than a
gate - and step 4 swaps the very tables it constrains, so the measurement has
to survive into that step or it protects nothing.
Comparison is exact-within-scope per spec section 4: every listed finding must
be raised and no other lexicon finding may be. Findings are named by commons
pattern_id, which scanForInjection() does not carry - it returns our labels.
Section 3.1 permits a runtime registered in the lexicon's aliases object to
compare through it, and we are registered. Measured first, not assumed: the
map is a total bijection, 83 labels to 83 ids, no duplicates, family membership
agreeing throughout. Nothing here restates a pattern's id, severity or label,
so nothing here can drift from the lexicon.
Deliberately NOT done: adding an id field to our 83 table entries. It would
change the source file the golden gate pins by sha256, forcing a re-bless in
the middle of a behaviour-preservation measurement, and duplicate what step 4
does anyway when the table itself starts loading from commons JSON.
Case discovery is driven by manifest.cases and cross-checked against the
directories on disk, because section 1 requires every case to run and a
deleted case dir would otherwise shrink the gate silently. An unimplemented
match or scope throws rather than skips (section 4). Input bytes and sha256 are
both verified before scanning - two fixtures carry characters invisible on
screen.
Proven red-capable in both directions by mutating the subject, not the harness:
neutering one pattern failed exactly override__disregard; widening one to
[aeiou] failed 82 cases on extra findings. Source restored byte-identical after
each.
Suite 2158, 85 new. The one red in the parallel run is the known
pre-compact-scan size-cap flake (366 ms alone, 1060 ms under load).
Vendoring commons v0.1.0 under scanners/commons/ falsified the premise of
`degrades gracefully when commons has not been vendored yet`: it asserted the
default DEFAULT_COMMONS_ROOT did not exist. It was the only red test after the
subtree add (2072/2073).
The graceful-empty contract it guarded is covered twice over by the
missing-artifact and invalid-JSON cases, which drive the same code path through
an explicit commonsRoot. So the replacement asserts the direction that is now
uncovered and matters more: a non-zero record count through the real default
root, no override.
That is the positive load-assertion Phase 5 step 4 requires. Every other gate we
have treats a commons load failure as indistinguishable from a legitimately
empty table, so a total loss of the vendored corpus would leave the suite green.
Proven red-capable by moving scanners/commons aside: the assertion fires by
name, not as an incidental TypeError elsewhere.
Suite 2073/2073.
Two of the three confirmed v8.x-A evasions are closed; the third (attribute
padding) stays open by operator decision and is documented, not silently left.
Recall fix. hybrid-xss required a closing </script>, so `<script>alert(1)` and
`<script src=x.js>` both passed scanForInjection() with found: false, while the
closed form returned high. A src= tag has no body to close in the first place.
The opening tag alone is the signal, and matching it is strictly linear: one
negated-class run whose excluded character is its own terminator, so there is
no backtracking surface that could re-introduce v7.8.3 #24.
ReDoS gate. The #24 test covered six html-obfuscation patterns against the two
shapes that defect was found on. A catastrophically backtracking regex added
anywhere else in the four tables would have failed no test at all - which is
why #24 had to be found by hand. The new gate times every exported pattern and
asserts its own coverage, so it cannot be narrowed silently.
It has two classes because the blowup shapes need opposite inputs, and because
a synchronous RegExp.test() cannot be interrupted: an exponential pattern met
with a 512KB input would HANG the run rather than fail it. Class 1 uses a
28-char ambiguous-run ladder, where exponential costs ~1s and anything sane
costs microseconds; on a hit the 512KB sweeps refuse to run. Class 2 uses the
hook's real read cap to catch the polynomial #24 class. Proven to fire: an
injected exponential regex was named by label and the hang guard held.
Measured clean: worst single pattern 17ms, full large sweep 218ms.
Golden gate went red as expected and was re-blessed - diff is exactly two
lines, the regex source and the source-file digest. The 61-payload reference
run is byte-identical, so no showcase behaviour moved.
Suite 2073/2073 (was 2063; +5 recall tests, +5 gate tests).
Still open, unchanged: all 7 bounded HTML patterns evade on >256 chars of
padding, in two positions - before the attribute and inside the style value.
Wider than STATE recorded, which named only aria-label.
Advisor review on the prior commit (69cad7c) caught a real detection-kill
vulnerability before push: reading `commons.root` from the SCANNED
TARGET's .llm-security/policy.json let a hostile cloned repo redirect
llm-security's own detection corpus to an attacker-supplied (empty)
one, with graceful-empty fallback making the substitution silent — a
substitutive override, unlike sig.custom_rules_path's additive one.
Dropped the policy import entirely; commons location is this plugin's
own concern, resolved only from __dirname or an explicit test/dev
override, never from policy or the scan target.
Also fixed two issues the review surfaced:
- Default vendor path was repo-root `shared/`, which package.json's
`files` allowlist (bin/, scanners/, knowledge/) would never publish —
moved under scanners/commons/, inside the directory that actually
ships. Same defect class as 2fe2915 (green dev checkout, empty
detection tables once installed).
- Cache keyed success/failure together, so the first caller's
`fallback` shape (e.g. []) leaked to a second caller expecting a
different shape ({}) on the same missing artifact. Cache now stores
a load-failed sentinel and returns each caller's own fallback.
Loaded artifacts are also deep-frozen, since the cache hands out one
shared object by reference to every caller.
New/changed tests cover all four: a simulated hostile-target policy
file is ignored, the failure-cache no longer cross-contaminates
fallback shapes, and mutating a loaded artifact throws.
Golden baseline unchanged; full suite 2063/2063.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QAYkRaBXT6tmWXTQAi1ZBg
Thin, sync-read JSON artifact loader for the future vendored
llm-security-commons subtree, modeled on signature-scanner.mjs's
loadRules()/loadCustomRules() pair: process-cached, graceful-empty
fallback on any read/parse error, and policy-extensible via a
`commons.root` policy value (mirrors sig.custom_rules_path).
Unit-tested now against a local fixture — Phase 4 (commons repo
creation, gated on the operator creating the Forgejo remote) hasn't
run yet, so the default `shared/` vendor path doesn't exist in this
checkout. That "not vendored yet" case is itself asserted: the loader
must degrade to the caller's fallback, not crash.
Not wired to any consumer yet (that's Phase 5 step 4, table-by-table
behind the golden gate). No CLI/hook/scanner-visible behaviour exists
to document. Golden baseline unchanged; full suite 2063/2063 (2053 +
10 new).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QAYkRaBXT6tmWXTQAi1ZBg
The `coverage` block never looked at the 61 hook invocations. It probed every
payload STRING against every injection regex in-process, so `47/83` meant
"if you threw all 61 strings at all 83 patterns, 47 would match" - not "the
reference run exercised 47". The pre-bash-destructive payloads never reach
injection-patterns at all, yet their strings were in the probe set and could
mark a pattern exercised.
That number was asserted as reference-run coverage in three places that
instruct a future session: tests/golden/README.md, the STATE golden-gate
section, and the generator's summary line. In a repo whose v7.8.2 lesson was
"the check reported success without running", a figure that measures one
thing while labelled another is the same defect wearing a different hat.
Relabelled rather than re-measured - the probe still honestly bounds the gate
(an unreachable pattern is one the corpus cannot protect under ANY
attribution), it just has to say what it is:
coverage.kind = 'static-reachability', with the caveat in a `note` field.
patternsExercised -> patternsReachable
uncoveredPatterns -> unreachablePatterns
tablesExercised -> corpusContains (a payload CONTAINS a homoglyph; it
does not say the run folded one)
The gate now pins both the `kind` and the note, so the honest label cannot be
dropped quietly by a later edit.
Also surfaced the gap the old wording hid: the four OWASP maps have NO
behavioural coverage - they are scanner-side and no hook in this corpus
reaches them. They are precisely the tables the dump was widened to cover, so
the table digest is their only protection. Stated in the README next to the
47/83 line, where a reader was previously left to infer the reference run
backed them.
Flakiness check for the new gate (61 sequential spawns, ~12s, the most
process-heavy file in the suite, added to a suite npm test runs concurrently):
three consecutive full runs, 2053/2053, 0 fail. The three known
timing-sensitive files did not destabilise.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BB4vXvwvtW4dxbPRd6vsez
Records the reference artifacts Phase 5 swaps will be measured against, and
the gate that reads them. No extraction yet: this is the "before" picture,
and it had to land first or every later comparison would be confounded.
Typed test/ rather than feat/ deliberately - the only production change is
one widened export; the rest is gate, artifacts and generator.
Written failing-first (7 red on missing artifacts), then generated.
Three layers, because the plan's "assert .source/.flags of every regex" is
necessary but not sufficient:
1. regex records - .source/.flags off the COMPILED object. After a swap a
pattern is new RegExp(jsonString, flags), so the plan's named hazard
(JSON backslash-doubling on 83+18 regexes) is visible here and nowhere
else. Source-text comparison cannot see it.
2. table records - key/value digests. Most of what Phase 4 moves is not a
regex at all: HOMOGLYPH_MAP (x3, AS-IS), the typosquat tokens and the
four OWASP maps are char->char and string->string data. A regex-only
dump is blind to a broken homoglyph swap, i.e. to the bulk of the
payload. Operator decision: widen the dump.
3. file records - sha256 of the five moving-set sources. This dissolves
STATE's open question (how to enumerate every regex): it is complete by
construction, covering inline regexes in function bodies that no export
walk reaches, with no JS parser in a zero-dep repo. A lexical count
would have pinned a lie - severity.mjs scores 4 "regexes" that way and
exports none.
Both layers were proven to fire, not assumed to: mutating one HOMOGLYPH_MAP
entry reddens the table layer, and widening an inline regex inside
decodeHexEscapes (unreachable by any export walk) reddens the file layer.
HOMOGLYPH_MAP is now exported from string-utils.mjs. That export is a source
change the plan already flags as a surface hazard ("private tables become
loaded"), so it is pre-paid here rather than confounding the before/after.
Reference run: the 61 showcase payloads through the real hook entry points,
sequentially - array order is semantic and the plan forbids key-sorting, so
concurrency is removed rather than sorted away. 61/61 match expectation,
which also settles the plan's open assumption that payloads.json expectations
match current behaviour. Coverage is recorded, not assumed: 47/83 patterns,
with the other 36 listed by key so the gate never implies coverage it lacks.
Suite counts are per-file, each file run alone - a total is unattributable,
and the three known timing-sensitive files flake only under concurrency.
2045 pass / 0 fail across 91 files, matching the pre-existing count exactly.
The gate's own file is excluded (it reads the artifact the run produces) and
that exclusion is named in the artifact.
Suite: 2053/2053, 0 fail.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BB4vXvwvtW4dxbPRd6vsez
Closes Phase 3 (B11) of the v8.0.0 plan. Three parts, all with the failing
test written first.
riskScoreV1 removed. scanners/lib/severity.mjs drops riskScoreV1() and its
SEVERITY_WEIGHTS_V1 table - @deprecated since v7.0.0, kept for diff/comparison,
zero callers in code or tests (re-verified, not taken from the plan). The v1
weights are recorded in CHANGELOG so an old score stays re-derivable. riskScore
(v2) is untouched; a test pins that one critical still lands in the 70-95 tier
and that 50 lows score below it, which is exactly the case v1 collapsed to 100.
Posture category 12 no longer keys off an identifier name. The check was
/TRIFECTA_MODE/i over the session-guard source, which measured what a constant
was CALLED rather than whether enforcement was configurable. With the env-var
gone, that regex would have dropped every correctly-migrated project from PASS
to PARTIAL - the gate punishing the migration it exists to encourage. It now
matches getPolicyValue('trifecta', 'mode', ...) and still accepts a pre-v8
vendored guard reading the old env-var, because a third-party project carries
its own hook copy and is equally configurable either way; the evidence line
says which of the two was found. The PARTIAL finding recommended setting an
env-var that v8 ignores; it now names the policy key. The grade-a fixture hook
moves to the policy-era form.
Two never-implemented env-vars deleted from the docs. LLM_SECURITY_SCR_OFFLINE
(ci-cd-guide) and LLM_SECURITY_OFFLINE (supply-chain-attack example) were
documented as OSV.dev / npm-audit kill-switches. No code has ever read either -
verified by grep across scanners, hooks and scripts, which finds them only in
markdown. A promised kill-switch that does nothing is worse than a documented
absence: it is trusted precisely when the run is meant to be air-gapped. The
docs now say there is none and that egress must be blocked at the network
layer. The LLM_SECURITY_AUDIT_* wildcard is narrowed to the one real key.
Docs. Migration section in README + CHANGELOG with the env-var -> policy-key
table, the detection commands (env + shell rc + .envrc + workflows), and the
explicit warning that a removed variable is now INERT rather than an error -
which is the failure mode that loses a project its configuration silently. The
hardening-guide env table splits into surviving vars and a removed-vars
migration table; its "promote to block" runbook named two variables that no
longer exist. Also swept: CLAUDE.md hook table, scanner-reference, ci-cd-guide,
both lethal-trifecta example docs, mitigation-matrix, injection-research.
Test counts in README/CLAUDE.md synced 2034 -> 2045.
Suite 2045 tests, 0 fail (2039 + 4 posture-trifecta + 2 riskScoreV1). The two
known parallel-load flakes did not recur this run.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BB4vXvwvtW4dxbPRd6vsez
BREAKING CHANGE: the four LLM_SECURITY_* configuration env-vars deprecated in
v7.3.0 are removed. .llm-security/policy.json is now the only source:
LLM_SECURITY_INJECTION_MODE -> injection.mode
LLM_SECURITY_TRIFECTA_MODE -> trifecta.mode
LLM_SECURITY_ESCALATION_WINDOW -> trifecta.escalation_window
LLM_SECURITY_AUDIT_LOG -> audit.log_path
LLM_SECURITY_DEPRECATION_QUIET -> dies with the mechanism it silenced
Setting a removed var is now inert - it does not warn, and it does not
configure. Env-vars with no policy equivalent (PRECOMPACT_MODE,
PRECOMPACT_MAX_BYTES, UPDATE_CHECK, MCP_CACHE_FILE, IDE_ROOTS) are unaffected.
getPolicyValueWithEnvWarn and its one-shot stderr warning are deleted from
policy-loader.mjs, along with the module-scoped warned-var Set. The four call
sites collapse to getPolicyValue. getPolicyValue's JSDoc claimed "environment
variables ALWAYS take precedence" - it never read env itself, so that line
described the shim, and it is corrected rather than deleted.
User-facing hook strings that advertised a removed var as the escape hatch now
name the policy key instead: the inject-scan block reason, its warn-mode note,
both escalation-window advisories, and the trifecta block message. A blocked
user following the old text would have set a var that does nothing.
Tests. tests/lib/v8-env-removal.test.mjs is the regression gate and was written
failing first (8 of 12 red before the change). It pins the NEGATIVE - setting a
removed var does not alter the outcome - because that is the half that rots
silently: a re-introduced process.env read would leave every migrated positive
test green, since those configure through policy.json and never set the var at
all. One assertion walks hooks/scripts and scanners for `process.env.<removed>`
so the re-introduction is caught structurally, not only behaviourally.
The 44 env-driven test occurrences (18 inject-scan, 13+4 session-guard, 9
audit-trail) migrate to a throwaway .llm-security/policy.json via a new
runHookWithPolicy helper in hook-helper.mjs; audit-trail runs in-process, so it
supplies the same policy through CLAUDE_PROJECT_ROOT. The D3 mechanism tests in
policy-loader.test.mjs are deleted with the mechanism.
Suite 2039 tests, 2037 pass (+12 gate, -7 D3 mechanism). The 2 failures are the
known parallel-load timing flakes (pre-compact-scan size-cap,
pre-install-supply-chain F-3); both green when run isolated.
Remaining in Phase 3: posture-scanner TRIFECTA_MODE heuristic, riskScoreV1
removal, ghost-var cleanup, docs + migration note.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BB4vXvwvtW4dxbPRd6vsez
repo-standard v0.2.0 flagged two README findings via coord. Verified both
against ground truth before acting:
- BADGE-COUNT (WARN, real): 8 badges confirmed by manual count. All eight
numbers were accurate -- commands 20, agents 6, hooks 9, knowledge 23,
scanners 22 (27 top-level .mjs minus the 5 non-scanner modules, per the
counting rule in docs/scanner-reference.md). The finding is about clutter,
not staleness. Operator chose to keep Version, Platform, Scanners, License
and drop the four inventory badges (Commands/Agents/Hooks/Knowledge) --
each of those repeats a table further down the same page and adds a sync
obligation to every version bump.
- README-H1 (WARN, operator's call): no change. The H1 is the human-readable
product title; llm-security is the package id in plugin.json and the
catalog. Renaming the H1 to the slug would be a readability regression.
doc-consistency.test.mjs pinned two of the removed/kept badges. The scanners
assertion still holds; the knowledge_docs badge assertion is dropped and the
test narrowed to the README prose claims, which remain derived from
knowledge/ contents. Full suite 2034/2034.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016h37aUdBLVDT9xrvG9osA9
Extends tests/lib/doc-consistency.test.mjs with 15 cases that derive every
inventory count from source instead of trusting prose. Each count has one
stated derivation; a doc surface that disagrees now fails the suite.
Counts corrected (all were wrong before the test existed):
- orchestrated scanners: docs said 10 (README, ci-cd-guide), CLAUDE.md said 12,
the synthesizer agent said 9 — scan-orchestrator registers 14
- total scanners: README badge + 3 prose sites said 23; the counting rule in
docs/scanner-reference.md (14 orchestrated + 8 standalone) yields 22
- knowledge files: README badge + prose said 22; knowledge/ holds 23
- output.mjs finding() prefix JSDoc listed 10 of the 17 prefixes actually
passed to it (missing IDE, MCI, MEM, PST, SCR, TFA, WFL)
- norwegian-context.md said "8 hooks, 10 scanners" -> 9 and 14
- ci-cd-guide "what gets scanned" table listed 10 of 14 rows; adds workflow,
trigger abuse, signature, AST taint
Two plan items changed after verifying against ground truth:
- CLAUDE.md's synthesizer "(12 scanners)" was not a deliberate subset; the
agent file itself claimed 9. Both bumped to 14.
- compliance-mapping.md's "13 posture categories" is substantively correct —
its matrix has exactly 13 data rows, and categories 14-16 are governance
consumers of the file, not rows in it. The planned 13->16 bump would have
made the document false. Wording clarified to "code-level" instead, and the
test now pins row count against the stated claim.
Framework currency (both verified against primary reporting):
- EU AI Act: Digital Omnibus (EP 2026-06-16, Council 2026-06-29) deferred the
high-risk obligations behind Art. 9/15 to 2027-12-02 (Annex III) and
2028-08-02 (Annex I); transparency still applies from 2026-08-02
- OWASP Agentic AI Top 10 labelled as the 2026 edition
Also: CLAUDE.md Distribution section rewritten monorepo -> polyrepo (each
plugin is its own repo; the catalog pins url + ref per plugin), and
current-state test counts synced 2013 -> 2034. Release-note paragraphs keep
their historical numbers.
No scanner, hook, or command behaviour changes. Suite 2034/2034.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wt4YQGoXwRja5K2Zmv8RZE
#20 git-forensics' reflog force-push detector had a redundant bare 'reset' term subsuming 'reset:', so any commit subject containing 'reset' tripped it; removed. #22 diff-engine's per-current moved-fallback greedily consumed a baseline candidate a later byte-exact match needed (mislabeling unchanged as new/moved on duplicate fingerprints); a global exact pass now runs before the moved-fallback. #54 memory-poisoning double-reported a 64+ char hex token as both base64 and hex; the base64 check now skips pure-hex tokens.
#56 SARIF output hardcoded driver.version 6.0.0 because the orchestrator called toSARIF() without a version; it now passes the real plugin version from package.json. #50 the VS Code known-malicious blocklist was empty with no explanation (unlike the JetBrains file's 'empty by design' note); added a matching blocklist_note. Suite 2004/0.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TcQyMTQfyrsAapaCMPxTtQ
#21 the bash normalizer decoded only \xHH, leaving ANSI-C octal/\u/\U forms literal so canonical rm/curl never surfaced; now decodes all three. #23 file-discovery keyed on extname so .env.local/.env.example (extname .local/.example) were silently skipped; now matches multi-part suffixes. #42 a legitimate leading UTF-8 BOM was flagged HIGH (and the tool's own auto-cleaner refused to strip it); pos-0 BOM now excepted. #52 collapseLetterSpacing used a literal space, letting multi-space/tab spacing evade; now [ \t]+. #55 redact(_,60,0) did slice(-0) and leaked the whole unredacted URL; showEnd===0 now means no tail.
#30 embedded base64 (const x = "<base64>") never satisfied the whole-string decode, so the SIG identity engine never saw it; added decodeEmbeddedBase64 as an OPT-IN param on normalizeForScan (default off — appending a decoded copy would double-count per-match findings, e.g. content-extractor's injection scan) and enabled it only in signature-scanner, which dedups variants. #36 signature-scanner ignored the documented sig.custom_rules_path policy option; now loads+merges custom rules through the same family filter. Suite 2004/0.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TcQyMTQfyrsAapaCMPxTtQ
#33 the frontmatter parser collected a block-scalar body (description: |) but did not skip it, so an indented name:/allowed_tools: inside the body re-matched as a top-level key and overrode the real values TRG-shadow and permission checks depend on; the parser now consumes block-scalar bodies as opaque content. #32 block-scalar headers carrying indentation/chomping indicators (|2, >-, |-2) were not recognized, so their bodies never reached the run: injection sink; now matched via a proper indicator/chomping regex.
#43 the B4 actor auth-bypass detector inspected only braced ${{ }} expressions, missing the canonical bare 'if: github.actor == ...' form (Synacktiv Dependabot-spoof false negative); bare if: expressions now emit a synthetic event the detector reads. Suite 2004/0.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TcQyMTQfyrsAapaCMPxTtQ
#24 the HTML-obfuscation injection patterns had overlapping unbounded runs plus a required closing quote, so a non-closing input backtracked O(N^2) (~28.7s at the 512KB cap); quantifiers bounded, pathological input now 4ms. #53 mcp-live-inspect buffered MCP-server stdout via readline with no cap, so a hostile stdio server could exhaust memory / throw an uncaught RangeError; replaced with manual line buffering capped at 4MB that rejects pending RPCs and destroys stdout. #31 vsix-fetch's same-host redirect follower had no depth cap (loop hang); added depth>=5 cap mirroring the sibling fetcher.
#25/#51 mcp-description-cache and skill-registry wrote JSON via bare writeFileSync (non-atomic: concurrent load-modify-save loses updates, a torn read silently yields an empty registry); both now write a temp file then renameSync. Suite 1931/0.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TcQyMTQfyrsAapaCMPxTtQ
#38 toxic-flow matched trifecta-leg keywords with bare includes(), so substrings ('url' in 'curl', 'key' in 'monkey', 'auth' in 'author') fabricated CRITICAL trifectas on benign components; now word-boundary matched. #40 TRG-broad fired HIGH on a bare any/all/every anywhere ('fix any lint errors'); the universal-claim regex now requires genuine universal phrasing. #41 TRG-baiting substring-matched ('any file' in 'many files'); now boundary-anchored. #39 the broad-name list missed multi-char generic names (helper/assistant/auto/general/agent/tool); widened coherently so it does not reintroduce #40. #57 the '(recovered from obfuscation)' label compared raw against a lowercased normal form, firing on any uppercase char; now gated on an explicit decode-changed flag.
#26 (same file) getPolicyValue used 'key in sectionObj' with no type guard, so a scalar section override in policy.json (e.g. {"injection":"block"}) threw an uncaught TypeError; now guarded to fall back to the default. Suite 1931/0.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TcQyMTQfyrsAapaCMPxTtQ
Wave C step C1: extend the MCP description cache schema with a sticky
baseline slot per tool and a rolling history array (last 10 drift events).
Cumulative drift = levenshtein(current, baseline) / max(|current|, |baseline|);
emits a separate signal when ratio >= mcp.cumulative_drift_threshold
(default 0.25). Per-update drift logic and threshold unchanged.
- loadCache(): TTL purge now skips entries with a baseline, preserving
cumulative-drift detection across the 7-day window. v7.2.0 entries
(no history field) are migrated on read by seeding baseline from the
current description and adding an empty history array. Entries with
history but no baseline (post-clearBaseline) are NOT re-seeded.
- checkDescriptionDrift(): when an entry exists with history but no
baseline (i.e. baseline was cleared), the next call re-seeds baseline
from the incoming description so the legitimate next version becomes
the new baseline.
- clearBaseline(toolName?): removes baseline for one tool or all tools.
Preserves description / firstSeen / lastSeen / history.
- listBaselines(): read-only listing for the upcoming reset CLI.
- LLM_SECURITY_MCP_CACHE_FILE env var override for end-to-end testing.
- New policy key mcp.cumulative_drift_threshold (default 0.25).
Tests: 23 new unit tests; existing 10 still pass.
Docs deferred: CLAUDE.md update lands in C3 alongside the new
/security mcp-baseline-reset command. C2 adds the hooks-table footer
note. Combined wave docs match plan §"Wave C — Touch" list.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds scanGitAttributes(repoDir) — pure function that parses
.gitattributes after a sandboxed clone and returns the
{filter,diff,merge} driver entries that would run on checkout. The
clone CLI prints each entry as a "MEDIUM" stderr advisory followed by
a recommendation to verify the smudge/clean command before moving the
clone outside the sandbox.
Why: filter drivers execute arbitrary shell during checkout (smudge
runs on read, clean on write). Even with the existing sandboxed clone,
downstream consumers that re-checkout files outside the sandbox can be
exploited. Surfacing the directive list lets the caller decide whether
to proceed.
Out-of-scope: in-line content of the smudge command is not analysed —
the advisory is for human review, not automatic blocking.
Tests:
- tests/lib/git-clone-gitattributes.test.mjs (8 cases): LFS-style,
custom driver, missing/empty/comment-only files, line-number
tracking, inline-comment stripping, unreadable path graceful return.
Adds rot13 to the variantSet built in scanForInjection(), so
imperative phrases hidden as rot13 inside code comments still hit
the existing CRITICAL/HIGH/MEDIUM pattern arrays.
normalizeForScan() already covers base64, hex, URL, and HTML decoding
in a 3-iteration loop — those are NOT duplicated here. rot13 is the
only genuinely new variant: it is its own inverse and not part of any
NIST/Unicode normalization spec, so it has to be applied explicitly.
Threshold: only inputs >40 chars enter the rot13 pass, to suppress
false positives on accidental letter-shifts in tokens, ids, and short
identifiers. Variants are deduplicated against the existing set so
matchers do not run twice.
3 new tests in injection-patterns.test.mjs (rot13 detection, sub-40
char suppression, plaintext path still green). Total 168 tests pass.
Closes E3 in critical-review-2026-04-20.md.
Critical-review §4 E16 finding: pre-v7.2.0 homoglyph normalization fired
ONLY for the MEDIUM-advisory "obfuscation present" signal. Pattern
matchers in scanForInjection compared against raw + decoded variants
only — they did NOT compare against a fold-normalized variant. As a
result, "ignоre previous instructions" (Cyrillic о, U+043E) bypassed
the CRITICAL "ignore previous" pattern.
Two coordinated edits:
scanners/lib/string-utils.mjs
- Adds HOMOGLYPH_MAP (frozen) — surgical Cyrillic/Greek → Latin map.
~25 entries focused on injection-vocabulary letters
(a, e, o, c, p, x, y, i, j, s, l, A, E, O, C, P, X, Y, T).
- Adds foldHomoglyphs(s) — pipeline: NFKC → apply HOMOGLYPH_MAP.
NFKC handles Mathematical Alphanumeric (U+1D400 block), fullwidth
Latin (U+FF21 block), ligatures, width variants.
Excluded by design from HOMOGLYPH_MAP:
- Latin Extended (æ, ø, å, é, è, ñ, ü, ö, ä, ç, ß, þ, ð) — legitimate
Norwegian/German/French/Spanish letters. Map them and we false-positive
on every non-English source file.
- Greek letters not visually overlapping (β, γ, δ, ...)
- Cyrillic letters not visually overlapping (б, г, д, ж, ...)
scanners/lib/injection-patterns.mjs
- scanForInjection now builds a 4-variant set: raw, normalized,
folded(raw), folded(normalized). Set deduplication skips redundant
identical variants. Existing dedup-by-label (seenLabels Set) prevents
double-counts when the same pattern matches in multiple variants.
- foldHomoglyphs added to the imports.
Tests: +27 cases in tests/lib/string-utils-homoglyph.test.mjs:
- 6 Cyrillic → Latin (lowercase, uppercase, multiple substitutions,
Palochka U+04CF)
- 3 Greek → Latin
- 2 NFKC normalization (Math Bold, Fullwidth)
- 8 preserves-non-confusable (Norwegian æøå, German umlauts, French
accents, Spanish ñ, emoji, CJK, Arabic/Hebrew)
- 3 edge cases (empty, null/undefined, idempotency)
- 5 scanForInjection integration (Cyrillic ignore, Cyrillic Assistant,
Norwegian non-trigger, benign "ignore" comment, mixed Cyrillic+Greek)
Test-development found: U+1D5DC is "I" not "A" (test pin caught my
codepoint mistake — fixed during dev).
Suite: 1617 → 1644 (+27). All green.
Critical-review §4 E1 finding: pre-v7.2.0 the Unicode-stego detector
(`containsUnicodeTags`) covered only U+E0001-E007F (Tag block). Private
Use Areas — also invisible in most terminals and surviving normalization
— were not detected. Attackers could encode payloads in PUA codepoints
that pass through `scanForInjection` undetected.
Coverage extended to:
- U+E0001-E007F Unicode Tag block (existing — DeepMind kat. 1)
- U+F0000-FFFFD Supplementary PUA-A (NEW — E1)
- U+100000-10FFFD Supplementary PUA-B (NEW — E1)
Detection-only for PUA: PUA characters have NO standard ASCII mapping,
so `decodeUnicodeTags` leaves them unchanged. Detection alone is
sufficient — `scanForInjection` emits HIGH on any presence, regardless
of decoded content.
Function name `containsUnicodeTags` preserved for back-compat. All
existing call sites (injection-patterns.mjs:259, etc.) work unchanged.
Semantically the function is now "containsHiddenUnicode".
Tests: +21 cases in tests/lib/string-utils-hidden-unicode.test.mjs:
- 5 Tag-block regression guards
- 4 PUA-A range cases (start, just-inside, end, buried-in-ASCII)
- 3 PUA-B range cases
- 5 boundary cases (gap U+E0080-EFFFF, U+10FFFE noncharacter, emoji,
CJK, Latin Extended — all must be FALSE)
- 4 decodeUnicodeTags passthrough cases (PUA-A unchanged, PUA-B
unchanged, Tag block still decodes, mixed Tag+PUA)
Suite: 1596 → 1617 (+21). All green.
Critical-review §2 B7 finding: pure Levenshtein <=2 misses the most common
modern typosquat pattern — popular-name + token-injection suffix. Examples:
lodash → lodash-utils (edit distance 6, not flagged pre-B7)
react → react-helper (edit distance 7, not flagged pre-B7)
express → express-wrapper (edit distance 8, not flagged pre-B7)
Three coordinated edits:
scanners/lib/string-utils.mjs
- Adds tokenize(name): string[] splits on -/_, lowercases
- Adds tokenOverlap(a, b): number intersection.size / min(|a|,|b|)
- Adds TYPOSQUAT_SUSPICIOUS_TOKENS frozen list of common typosquat
suffixes. Excludes language-extension tokens (js, jsx, ts, tsx) — the
v7.0.0 allowlist contains `tsx` as a legit package and including the
same token in the suspicious set creates a contradiction. Caught by
the new allowlist-intersection-guard test. Also excludes 'pro'
(legitimate edition marker).
scanners/dep-auditor.mjs + scanners/supply-chain-recheck.mjs
- New checkTyposquatTokenOverlap() helper — fires AFTER Levenshtein 1/2
branches, only when:
1. popular package's tokens ⊆ declared name's tokens (strict superset)
2. declared name has at least one suspicious suffix
3. popular package is in topCutoff window
All three conditions required — conservative by design. Allowlist
precedence preserved (existing 22 npm + 13 PyPI entries always pass).
MEDIUM severity, NOT block. New finding title prefix:
"Possible typosquatting via token-overlap".
Tests: +21 cases across two new files
- tests/lib/string-utils-tokens.test.mjs (15) — tokenize, tokenOverlap,
TYPOSQUAT_SUSPICIOUS_TOKENS frozen contract, allowlist-intersection
guard (caught the tsx conflict on first run)
- tests/scanners/dep-token-overlap.test.mjs (7) — integration via
in-memory tmpdir fixtures: lodash-utils flagged, react-helper flagged,
express-wrapper flagged, lodash exact NOT flagged, allowlist tools
(knip/tsx/nx/rimraf) NOT flagged, react-router-dom (no suspicious
suffix) NOT flagged, react itself (equal token set, not superset)
NOT flagged.
Existing dep.test.mjs and supply-chain-recheck.test.mjs unchanged —
all green (149 → 149 regression guard).
Suite: 1570 → 1591 (+21). All green.
Critical-review §2 B3 finding: `riskScore({info: N}) = 0` silently masks
info-volume findings. The behavior was correct (info is scoring-inert by
design) but undocumented. Operators reading a report with N info findings
had no way to know they contribute zero to verdict/band.
Three coordinated edits:
- scanners/lib/severity.mjs JSDoc — explicit "Info severity" subsection
spelling out: scoring-inert, surfaced in owaspCategorize aggregates,
treat as observability telemetry not verdict input. @param updated to
mark info as accepted but ignored.
- CLAUDE.md v7.0.0 risk-score-v2 line — one-sentence anchor pointing to
severity.mjs JSDoc.
- tests/lib/severity.test.mjs — anchor test alongside the existing
4-critical=93 anchor: asserts riskScore({info: 50}) === 0,
riskScore({info: 1000}) === 0, verdict({info: 100}) === 'ALLOW',
riskBand(riskScore({info: 500})) === 'Low'.
Decision: skip the optional `infoScore()` helper from the brief. No
current consumer would use it; doc-only fix keeps API surface minimal.
Revisit if a consumer emerges.
Tests: 1522 → 1523 (+1 anchor block, 4 assertions). All green.
Closes A2 of v7.1.0 critical-review patch (docs/critical-review-2026-04-20.md):
- B4 (severity JSDoc): 4 critical = 93, not 90. Fixed in scanners/lib/severity.mjs:23
and CHANGELOG.md v7.0.0 tier description. The actual computation has always been
93 (70 + log2(5)*10 = 93.22 → round); only the docs were wrong.
- §5.4 co-monotonicity: new sweep test in tests/lib/severity.test.mjs over 15
representative count vectors. Asserts that (verdict, riskBand) agree under the
v7.0.0 contract for every case — catches future drift between riskScore tiers,
verdict cutoffs, and riskBand cutoffs. Includes a B4 anchor test (riskScore
{critical: 4} === 93) so doc/code drift fails loudly.
- B8 (CaMeL claims toned down): post-session-guard.mjs:646 comment block and
CLAUDE.md:184 Defense Philosophy bullet now describe the implementation
honestly — opportunistic byte-matching of truncated output fingerprints
(first 200 bytes, SHA-256/16-hex), not semantic data-flow tracking.
Trivially bypassed by mutation, summarisation, or re-encoding. Inspired by
CaMeL (DeepMind 2025), but not a CaMeL capability-tracking implementation.
Tests: 1495 → 1511 (+16: 15 sweep cases + 1 B4 anchor). All green.
Final commit in the trustworthy-scoring series. Bundles verdict cutoff
alignment, the last suite of tests, and all documentation touch-points
that quote version numbers or describe v7.0.0 behaviour.
Verdict/band co-monotonicity
- `scanners/lib/severity.mjs` — verdict cutoffs moved from 61/21 to 65/15
so `BLOCK >= 65`, `WARNING >= 15` locks onto the v2 riskBand() boundaries.
Prevents "BLOCK / Medium band" contradictions under the v2 formula.
Scanner hardening (bug fixes from v7.0.0 testing)
- `scanners/entropy-scanner.mjs` — `policy_source` now uses
`existsSync('.llm-security/policy.json')` instead of value-based check.
Old heuristic always reported 'policy.json' because DEFAULT_POLICY now
carries an `entropy.thresholds` section.
- `scanners/lib/file-discovery.mjs` — `.sass` and GPU shader extensions
(`.glsl, .frag, .vert, .shader, .wgsl`) added to TEXT_EXTENSIONS. Without
this, shader files were invisible to file-discovery, so they were never
counted as skipped by the entropy-scanner extension filter.
Tests
- `tests/scanners/entropy-context.test.mjs` (new, 24 tests) — A. File-ext
skip (4), B. Line-level rules 11-17 (8), C. Policy overrides (3).
Fixtures generate 80-char base64 payloads at runtime via
`crypto.randomBytes` to dodge the plugin's own pre-edit credential hook
on the test source.
- `tests/lib/severity.test.mjs` — rewritten with v2 scoring table (70
tests total, was 52).
- `tests/lib/output.test.mjs:243` — "1 critical = score 80" under v2
(was 25 under v1).
- Full suite: 1485/1485 green (was 1461).
Docs
- `CHANGELOG.md` — v7.0.0 entry with BREAKING CHANGES section.
- `README.md` (plugin + marketplace root) — version badge, history table,
plugin-card version string, test count.
- `CLAUDE.md` — header version, "v7.0.0 — Trustworthy scoring" summary
paragraph at the top.
- `docs/security-hardening-guide.md` — new section 6 "Calibration & false
positives" documenting v2 formula, context-aware entropy scanner,
typosquat allowlist, and §6.4 tuning workflow. Existing "Recommended
baseline" section renumbered to §7.
Version bump
- `6.6.0 -> 7.0.0` across package.json, .claude-plugin/plugin.json,
scanners/ide-extension-scanner.mjs VERSION const, README badge,
CLAUDE.md header, marketplace root README card.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Pre-installation verification of VS Code extensions via URL — fetch a remote
VSIX, extract it in a hardened sandbox, and run the existing IDE scanner
pipeline against it. No npm dependencies.
Sources:
- VS Code Marketplace (publisher.gallery.vsassets.io direct download)
- OpenVSX (open-vsx.org official API)
- Direct .vsix HTTPS URLs
Defenses:
- HTTPS-only, TLS verified, manual redirect with per-source host whitelist
- 30s total timeout via AbortController
- 50MB compressed cap, 500MB uncompressed, 100x expansion ratio
- Zero-dep ZIP extractor: zip-slip, absolute paths, drive letters, NUL bytes,
symlinks (Unix mode 0xA000), depth limits, ZIP64 rejected, encrypted rejected
- SHA-256 streamed during fetch, surfaced in meta.source
- Temp dir cleanup in all paths (try/finally)
Files:
- scanners/lib/vsix-fetch.mjs (HTTPS fetcher, host whitelist, streaming SHA-256)
- scanners/lib/zip-extract.mjs (zero-dep parser with hardening caps)
- knowledge/marketplace-api-notes.md (endpoint reference)
- 3 test files (48 tests added: vsix-fetch, zip-extract, ide-extension-url)
Tests: 1296 → 1344 (all green).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add threshold-based exit codes (--fail-on <severity>) and compact
output mode (--compact) to scan-orchestrator and CLI. Pipeline
templates for GitHub Actions, Azure DevOps, GitLab CI with SARIF
upload. CI/CD guide with Schrems II/NSM compliance documentation.
npm publish preparation (files whitelist, .npmignore). Policy ci
section for distributable CI defaults. Version 6.1.0.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
New policy-loader.mjs reads .llm-security/policy.json with deep-merge against
defaults that exactly match existing hardcoded values. Integrated into all 7 hooks:
- pre-prompt-inject-scan: injection.mode (env var still takes precedence)
- post-session-guard: trifecta.mode, window_size, long_horizon_window
- pre-edit-secrets: secrets.additional_patterns
- pre-bash-destructive: destructive.additional_blocked
- pre-write-pathguard: pathguard.additional_protected
- pre-install-supply-chain: supply_chain.additional_blocked_packages
- post-mcp-verify: mcp.volume_threshold_bytes, mcp.trusted_servers
Backward compatible: no policy file = identical behavior to v5.1.0.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
New audit-trail.mjs writes structured events to LLM_SECURITY_AUDIT_LOG path.
Integrated into post-session-guard at 6 warning emission points: trifecta,
escalation-after-input, data flow, volume threshold, slow-burn, behavioral drift.
No-op when env var not set — zero overhead for existing users.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add /ultraresearch-local for structured research combining local codebase
analysis with external knowledge via parallel agent swarms. Produces research
briefs with triangulation, confidence ratings, and source quality assessment.
New command: /ultraresearch-local with modes --quick, --local, --external, --fg.
New agents: research-orchestrator (opus), docs-researcher, community-researcher,
security-researcher, contrarian-researcher, gemini-bridge (all sonnet).
New template: research-brief-template.md.
Integration: --research flag in /ultraplan-local accepts pre-built research
briefs (up to 3), enriches the interview and exploration phases. Planning
orchestrator cross-references brief findings during synthesis.
Design principle: Context Engineering — right information to right agent at
right time. Research briefs are structured artifacts in the pipeline:
ultraresearch → brief → ultraplan --research → plan → ultraexecute.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
