Skip to content

Releases

v0.1.0.dev3 (pre-alpha): Basic/Bearer/NTLM rules, narrower KEY=value filter

Choose a tag to compare

@keith-nielsen keith-nielsen released this 18 Oct 09:09

Warning

Pre-alpha. Interfaces, file formats, the audit record format and defaults will change without
notice or migration. Do not rely on PPE to protect real personal data or secrets yet: the proxy
observes traffic, it does not mask it.

What's new since v0.1.0.dev2

Secret rules v3

  • HTTP authorization credentials. Three new rules, with or without an Authorization: name before them, each with a value check:

    • Basic: base64 of a printable user:password;
    • Bearer: a 16+ character token with letters and digits;
    • NTLM/Negotiate: the NTLMSSP or SPNEGO prefix.

    On Samsung CredData these cover 689 of 692 Basic, 289 of 289 Bearer and 108 of 126 NTLM credentials.

  • A narrower placeholder filter for KEY=value. v0.1.0.dev2 used Gitleaks' 1,448 stop words, which also dropped real passwords containing ordinary words (44 on Nemotron). A value is now dropped only if it:

    • contains a placeholder word (change, example, your, …) or a word naming a kind of credential (token, passw, secret, …);
    • is a template (${X}, {{x}}, <x>);
    • has three or fewer distinct characters;
    • or is unquoted code (os.environ[, get_token(...)).

Measured. CredData, without the one test-vector file that holds 18% of the credentials:

found flags that are credentials
v0.1.0.dev2 0.441 0.664
v0.1.0.dev3 0.516 0.686
Gitleaks 8.30.1 0.338 0.762
detect-secrets 1.5.0 0.370 0.342
  • Nemotron passwords found: 257 (v0.1.0.dev2: 215).
  • Known costs: Password123-style values (2 on Nemotron) and unquoted passwords containing a bracket (2).
  • Details: bench/2026-10-17-creddata.md.

Checks at this tag: ruff, strict mypy, 346 passed, 1 skipped, 21 xfailed; ppe verify 10/10; mutation check 52/52.

Not yet

  • Generic secrets without a key word or format are still mostly missed (3,013 of the credentials PPE misses on CredData).
  • No decoding of base64, hex or percent-encoded text before matching; no masking on traffic (phase 4).

Try it

uv venv --python 3.12 .venv && uv pip install --python .venv/bin/python -e ".[dev]"
.venv/bin/pytest && .venv/bin/ppe verify
echo 'curl -H "Authorization: Basic ZGVwbG95Oms5THcycXpY" https://api.example.test' | .venv/bin/ppe scan

v0.1.0.dev2 (pre-alpha): Gitleaks secret rules, CredData benchmark

Choose a tag to compare

@keith-nielsen keith-nielsen released this 17 Oct 22:03

Warning

Pre-alpha. Interfaces, file formats, the audit record format and defaults will change without
notice or migration. Do not rely on PPE to protect real personal data or secrets yet: the proxy
observes traffic, it does not mask it.

What's new since v0.1.0.dev1

Secret detection

  • Gitleaks 8.30.1's rules in Stage 1. The default rule file is vendored unchanged (MIT, credited in NOTICE), pinned by SHA-256, and applied the way Gitleaks applies it: keyword prefilter, secret group, entropy threshold, allow-lists, generic dedupe. 217 rules load; 5 need a file path and are skipped.
  • RE2 (google-re2) is a new required dependency. The rules run with Go regex semantics, unmodified, in linear time on any input.
  • PPE's own KEY=value rule now drops values containing a Gitleaks placeholder word (changeme, example, …).

Measured on real code: Samsung CredData (11,030 files, 15,753 labelled credentials; built with its own script, pinned by manifest hash). Without the one test-vector file that holds 18% of the credentials:

found flags that are credentials
PPE v0.1.0.dev1 0.325 0.449
PPE v0.1.0.dev2 0.441 0.664
Gitleaks 8.30.1 0.338 0.762
detect-secrets 1.5.0 0.370 0.342
  • Nemotron-PII secret recall 0.062 → 0.131; Gretel 0.178 → 0.293.
  • Cost: the placeholder filter drops 44 real Nemotron passwords that contain dictionary words.
  • Details: bench/2026-10-17-creddata.md.

Also

  • Earlier external benchmarks (Nemotron-PII, Gretel finance, Presidio baseline), with detector fixes found on them: IPv4 at a sentence end, IBAN followed by a word, non-ASCII e-mail, IPv6 inside a word.
  • The 2026-10-16 product review's "now" fixes.

Checks at this tag: ruff, strict mypy, 319 passed, 1 skipped, 22 xfailed; ppe verify 10/10; mutation check 47/47.

Not yet

  • Generic secrets without a key word or format are still mostly missed (3,810 of the credentials PPE misses on CredData).
  • No decoding of base64, hex or percent-encoded text before matching.
  • No masking on traffic (phase 4), span models or arbiter.

Try it

uv venv --python 3.12 .venv && uv pip install --python .venv/bin/python -e ".[dev]"
.venv/bin/pytest && .venv/bin/ppe verify
echo 'export API_KEY="Zq8vT2mX9pL4wR7nK3bY"' | .venv/bin/ppe scan

v0.1.0.dev1 (pre-alpha): shadow-mode proxy

Choose a tag to compare

@keith-nielsen keith-nielsen released this 15 Oct 19:48

Warning

Pre-alpha. Interfaces, file formats, the audit record format and defaults will change without
notice or migration. Do not rely on PPE to protect real personal data or secrets yet: the proxy in
this release observes traffic, it does not mask it.

First tagged snapshot of Prompt Privacy Enhancement: a testable privacy harness for LLM traffic.

What's in it

Phase 1 core

  • Stage 0 shadow text (folds full-width digits, zero-width and bidi characters) and Stage 1 rules:
    12 secret rules, NRIC/FIN (incl. M-series), cards, IBAN, email, phone, IPv4/IPv6, operator dictionary.
  • Surrogate swap and restore with format transfer, stable per conversation, unlinkable across
    conversations.
  • Fully sealed 2-of-2 audit log: hybrid ML-KEM-768 + X25519 HPKE, padded records, hash chain.
  • Synthetic labelled corpus, benchmark, ppe scan | mask | corpus | bench | explain | verify | audit.

Since phase 1

  • Stage 1 detection is now linear in input size: 0.80 / 1.70 / 5.1 ms/KB at 10 / 100 / 400 KB
    became 0.68 / 0.70 / 0.76 ms/KB, with byte-identical output.
  • ppe serve (phase 2a), a shadow-mode proxy for the Anthropic Messages API on aiohttp:
    • loopback only; forwards requests and responses unchanged (bodies byte for byte, SSE as it
      arrives)
    • records sealed observe events with keyed per-identifier digests, deduplicated per conversation
    • observation never delays or fails traffic
    • refuses to start with aiohttp's pure-Python HTTP parser
  • Adversarial test pass:
    • 27 ambiguous or malformed requests are refused with 400 and nothing is forwarded
    • accepted requests are re-framed cleanly
    • also covered: disconnects, stalls, size limits, concurrency, duplicate headers
    • 300-case mutation fuzz
  • Design: ADR-0005 amendment (measured span-model latency), ADR-0013 (restricted zone and sticky
    labels, draft), MAS AI risk management guidelines mapped to controls.

Checks at this tag: ruff, strict mypy, 130 tests, ppe verify 10/10.

Not yet

  • ppe serve has not been validated with a real client against a real provider.
  • Reply-flow observation, OpenAI wire formats, exposure ledger, sensor API.
  • Masking on traffic (phase 4), span models, arbiter, signed checkpoints, retention/shredding,
    break-glass broker.

Try it

uv venv --python 3.12 .venv && uv pip install --python .venv/bin/python -e ".[dev]"
.venv/bin/pytest && .venv/bin/ppe verify
.venv/bin/ppe audit keygen --dev-token && .venv/bin/ppe serve   # 127.0.0.1:8787 -> api.anthropic.com