argus-redact 0.7.13

Encrypt PII, not meaning. Locally.

argus-redact

English · 中文说明

Encrypt PII, not meaning. Locally.

The privacy layer between you and AI. Your identity stays on your device — AI gets the meaning, not you.

Rated PRvL-Gold on the PRvL reference suite — see the spec for what it measures.

from argus_redact import redact

redacted, key = redact("张三的电话是13812345678，身份证号110101199003074610", names=["张三"], lang="zh", salt=42)
print(redacted)
# expected: P-83811的电话是138****5678，身份证号ID-03292

print(sorted(key.items()))
# expected: [('138****5678', '13812345678'), ('ID-03292', '110101199003074610'), ('P-83811', '张三')]

pip install argus-redact

Three Promises

	Promise	How
🛡️	Protected — your PII never leaves your device	3-layer local detection: regex → NER → local LLM
🧠	Usable — AI can still understand and help you	Pseudonym replacement preserves meaning and context
🔄	Reversible — substring-level inverse via per-message key	One-line restore() for verbatim LLM echoes; paraphrase / coref handled by compose layer, best-effort

Other tools shred your PII — it's gone forever. argus-redact encrypts it with a different key every time. ETH Zurich research shows LLMs can deanonymize users for $1-4/person when pseudonyms are fixed. We generate fresh random keys per call — the cloud sees unrelated pseudonyms every time.

Default redaction output

redact() emits per-type pseudonym codes, not Chinese label literals:

>>> redact("员工张三，身份证110101199003074610，电话[phone]", mode='fast', lang='zh')
('员工P-83811，身份证ID-89732，电话138****5678',
 {'P-83811': '张三', 'ID-89732': '110101199003074610', '138****5678': '13812345678'})

Type group	Default output	Strategy	Reversible
person / organization	P-NNNNN / O-NNNNN	pseudonym	✓
phone / email / bank_card	138****5678 (partial digits visible)	mask	✗
id_number / medical / ssn / ...	ID-NNNNN / MED-NNNNN / SSN-NNNNN	remove → per-type code	✓
self_reference	我 / 我妈 (kept verbatim)	keep	✓

To unify all reversible types under one prefix (hides PII type from the LLM):

redact(
    text,
    unified_prefix="R",
    config={
        "phone": {"strategy": "remove"},   # mask types must opt in to participate
        "email": {"strategy": "remove"},
    },
)
# → "员工R-83811，身份证R-89732，电话R-12345"

<TYPE_N> 1-based sequential token style is on the future-release candidate list (no committed timeline). See docs/configuration.md for the current strategy reference.

Privacy Levels

argus-redact evaluates your text from your perspective, not a regulator's:

🟢 Safe      — nothing about you is exposed
🟡 Caution   — contains personal info, not dangerous alone
🟠 Danger    — can narrow down to you specifically
🔴 Exposed   — directly identifies you

from argus_redact import redact

report = redact("身份证110101199003074610，手机[phone]，确诊糖尿病", report=True)
report.risk.level    # "critical"
report.risk.score    # 1.0
report.risk.reasons  # ("id_number (critical)", "phone (high)", "medical (critical)", ...)

This is what compliance frameworks don't tell you: how dangerous is it to share this specific text with AI?

Three Layers, Collaborative

Layer 1  Rust+Regex   phone, ID, bank card, email, self-reference, ...    <0.2ms
             │
         produce_hints() → text_intent, pii_density, self_reference_tier
             │
Layer 2  NER ← hints   locations, organizations, standalone names         10-100ms
Layer 3  Local LLM      implicit PII — symptoms→disease, behavior→belief  ~20s

Layers are not independent — L1 passes hints to L2, enabling collaborative detection. Instruction text ("帮我看看这段代码") skips NER entirely. High PII density lowers NER thresholds. Cross-layer agreement boosts confidence.

Unicode-hardened: NFKC normalization, zero-width stripping, Cyrillic/Greek confusable defense, Chinese digit detection (一三八零零一三八零零零 → detected as phone).

Core engine (regex matching, entity merging, restore, pseudonym generation) is written in Rust via PyO3 for maximum performance. Python handles orchestration, NER models, and LLM integration.

60+ PII types across 3 layers — from phone numbers to medical diagnoses, religious beliefs, political opinions. Default is mode="fast" (Layer 1 only, zero deps, sub-ms). Opt in: mode="ner" (+ NER models) → mode="auto" (all three layers).

Telemetry: ARGUS_PERF_LOG=perf.jsonl for per-call timing breakdown. Details →

Deployment fit — modes have very different latency budgets; pick by where you sit in the request path:

Mode	Latency (per doc)	Suitable as
fast	<1ms	Inline gateway plugin / hot LLM proxy path
ner	10–100ms	Sidecar / pre-flight middleware
auto	~20s (LLM-bound)	Async batch / offline review queue

Don't put auto in front of an interactive LLM call. Use fast inline + auto in a parallel audit lane.

Limitations & When NOT to Rely on This

argus-redact is a PII data minimization aid, not an anonymization or compliance certification:

• L1 fast (regex) matches well-defined formats. Novel or obfuscated variants, cross-field inference attacks pass through.
• L2 NER is statistical inference; out-of-distribution text (informal, typo-heavy, minority names) has higher miss rate. See benchmark results for measured numbers.
• No guarantee against adversarial inputs — attackers can craft text that evades detection.
• Removing explicit PII ≠ anonymity. LLM agents can re-identify individuals by combining residual, individually-non-identifying cues with public data — even on redacted text, even during benign tasks (Ko et al. 2026). Reversible substitution protects explicit identifiers and preserves LLM utility; it does not defend against inference-based re-identification, which a per-document redactor cannot fully prevent — the residual comes from combinations of quasi-identifiers, not single fields (why coarsening one field doesn't fix it).
• Not a GDPR / PIPL anonymization framework — anonymization is a compliance process decision, not a single-library output.

When to use argus-redact: reversible pseudonymization for LLM pipelines where you need redact() → LLM → restore() with zero PII crossing the network boundary.

When to consider alternatives: if you need one-way English PII masking with a single model call, OpenAI Privacy Filter and similar model-based maskers may fit better. argus-redact's strongest suit is reversible pseudonymization with per-message keys; Chinese has the deepest support (HanLP + native validators), the other 7 languages have regex + spaCy NER coverage. Pick by the workload, not by exclusivity.

Combine argus-redact with audit logging, rate limiting, and upstream policy — no single layer is sufficient.

8 Languages

	zh	en	ja	ko	de	uk	in	br
Phone	✓	✓	✓	✓	✓	✓	✓	✓
National ID	MOD11-2 + 15位旧版	SSN	My Number	RRN	Tax ID	NINO	Aadhaar	CPF/CNPJ
Bank/Card	Luhn	Luhn	—	—	IBAN	—	PAN	—
Person names	HanLP	spaCy	spaCy	spaCy	spaCy	spaCy	spaCy	spaCy
Email	✓	✓	✓	✓	✓	✓	✓	✓

Mix freely: lang=["zh", "en", "de"]. Pass known names: names=["王一", "张三"].

Performance

Rust core (PyO3), mode="fast" — p50, Apple M-series, Python 3.11. Reproduce with python tests/benchmark/bench_l1_rust_vs_python.py:

Text	redact()	restore()	Throughput
Short (17 chars)	0.03ms	<0.01ms	~29,000 docs/sec
Medium (770 chars)	0.75ms	0.15ms	~1,330 docs/sec
Long (10K chars)	9.3ms	0.18ms	~107 docs/sec

Pre-built wheels for all major platforms — no Rust toolchain needed to install:

✓ Linux x86_64 (glibc + musl/Alpine)
✓ Linux aarch64 (Raspberry Pi + Alpine ARM)
✓ macOS (Apple Silicon + Intel)
✓ Windows x64
× Python 3.10 / 3.11 / 3.12 / 3.13

Detection accuracy

Mode	Precision	Recall	F1
fast (regex)	81.6%	31.9%	45.8%
ner (+ spaCy)	74.9%	42.8%	54.4%
auto (+ Ollama 32B)	skipped this run

ai4privacy en, 500 samples, v0.7.9. auto mode skipped on the maintainer's hardware — see benchmark-report.md for full matrix + reproduction commands.

For context: fast mode is high-precision / low-recall by design — it only emits formats it can validate (Luhn, MOD11-2, etc.). Recall comes from ner and auto at the cost of latency. Pick the mode for your deployment shape (see Deployment fit above). Full benchmarks → | Performance →

North Star

Dimension	Current (v0.7.13)	Next milestone
Protected	60+ PII types, L1-L3. 0% PII leak on default profile across GPT-5 / Claude-Opus-4.5 / Gemini-2.5-Pro / GLM-4.5 in the PRvL reference suite. pseudonym-llm profile: 100% on three of four models; 96% / Bronze on Claude-Opus-4.5 (single reroll cell). Not a guarantee against adversarial inputs — see prvl-standard.md for full matrix. Cross-layer hints in 8 langs (zh/en/ja/ko/de/uk/in/br). SHAKE-256 derivation + full-salt entropy + faker identity-pass guard. State export omits salt by default; HTTP server refuses no-auth start; CLI writes O_NOFOLLOW + key files mode 0600; MCP token store TTL+LRU (v0.6.2). Windows CI + property-tested invariants + mutation-tested core (v0.6.3) + perf budget CI gate (v0.6.4) + session-isolation in integrations (v0.6.6) + README pinned-to-doctest + version-sync CI guard (v0.6.6) + compose namespace + pure-layer purity guard (v0.6.7) + seed→salt API rename + PIITypeDef SSOT + Presidio bridge through public redact + 3 new types (v0.6.8) + compose helpers shipped (v0.6.9) + Layer 1 freeze guards + KDF replay vectors + dead code subtract + manylinux digest pin (v0.6.10) + Adapter authoring surface (compose.register_pii_type / PIITypeDef / PatternMatch) + KDF replay edge cases (full-FF salt fix) + Layer 2 signature snapshot (v0.6.11) + HK/Macao travel permits + housing-fund zh L1 coverage (v0.6.12). v0.7.x — 100% Rust core SSOT: argus-redact-core crate + crates.io publish, with patterns/validators/normalization/replace+restore/fakers/person-scoring + the full L1 redact/restore engine ported to Rust (v0.7.0–v0.7.8) + fail-closed hardening & detection-correctness (v0.7.9–v0.7.10) + in-browser wasm build (v0.7.11). v0.7.12 — quasi-identifier detection breadth: evidence-gated zh bare-region, occupation, medical condition/allergy, and hobby (new type) detection via a shared evidence_detector framework, plus a re-identification eval (PRvL+ X-axis); the unreleased generalize strategy removed	Adversarial testing
Usable	PRvL U=100%. Pseudonym codes + realistic mode (zh + en + RFC shared) + per-call strategy overrides + keep strategy (whitelisted) + resumable streaming sessions + incremental streaming default + cross-language alias restore (zh ↔ en)	Task-aware guidance
Reversible	PRvL R by task: reference 100%, extract 50%, creative 0% (by design). Cross-language LLM rewrites (张三 → Zhang San) auto-restored via result.aliases + restore(text, key, aliases=...)	Task-aware guidance
Compliance	Meets PIPL Art.28 sensitive PII categories, risk assessment + profiles	PIPL/GDPR/HIPAA (byproduct)
Coverage	8 langs, 4 LLMs benchmarked, 6 frameworks	Browser extension

Risk Assessment

# Assess risk before sending to AI
report = redact(text, report=True)
report.risk.level         # "critical"
report.risk.pipl_articles # ("PIPL Art.28", "PIPL Art.51", ...)
report.entities           # detected PII details
report.stats              # per-layer timing

# CLI
argus-redact assess <<< "身份证110101199003074610"

Compliance profiles: redact(text, profile="pipl") / "gdpr" / "hipaa". Type filtering: redact(text, types=["phone", "id_number"]) / types_exclude=["address"].

Realistic Redaction (pseudonym-llm profile)

Default redaction emits placeholder labels ([TEL-79329], P-164) — clear for audit, but breaks downstream LLM reasoning because the message structure is gone. The pseudonym-llm profile replaces PII with realistic-looking but reserved-range fake values (e.g., 19999... mobile, 999... ID, 999999... bank card). LLMs reason correctly; humans can still tell it's synthetic if they know the convention.

Each call returns three text forms sharing one key dict:

Form	Example	Use for
audit_text	请拨打 [TEL-79329] 联系 P-164	Compliance archive — placeholder labels are auditable
downstream_text	请拨打 19999123456 联系张明	LLM input — semantic structure preserved
display_text	请拨打 19999123456ⓕ 联系张明ⓕ	UI rendering — visible ⓕ marker prevents confusion

from argus_redact import redact_pseudonym_llm, restore

# Chinese
zh = redact_pseudonym_llm("请拨打 13912345678 联系王建国", lang="zh")
zh.downstream_text  # "请拨打 19999123456 联系张明"           → LLM
zh.display_text     # "请拨打 19999123456ⓕ 联系张明ⓕ"        → UI

# English
en = redact_pseudonym_llm("Call [phone], SSN 123-45-6789", lang="en")
en.downstream_text  # "Call [phone], SSN 999-37-2811" → LLM
en.audit_text       # "Call [PHONE-23801], SSN [SSN-15772]"  → audit

# Mixed (auto-detect)
mx = redact_pseudonym_llm("客户Wang at [e-mail]", lang="auto")

# Round-trip works on any of the three forms, in any language
restore(zh.downstream_text, zh.key)   # → original
restore(en.downstream_text, en.key)   # → original
restore(mx.downstream_text, mx.key)   # → original

# CLI emits all three forms as JSON
echo "Call [phone]" | \
  argus-redact redact -k key.json --profile pseudonym-llm -l en | \
  jq .downstream_text
# "Call [phone]"

Reserved ranges:

• zh: 199-99-XXXXXX mobile (sub-segment unassigned by 工信部), 099- landline (no such area code), 999XXX ID address code (GB/T 2260 unassigned), 999999 bank BIN (银联 unassigned), 滨海市 fictional city.
• en: (555) 555-01XX phone (FCC permanent fictional reservation), 999-XX-XXXX SSN (SSA never assigns 9XX), 999999 credit card BIN, John Doe / Jane Roe person, 1313 Mockingbird Lane address.
• shared (RFC): example.com / .org / .net email (RFC 2606), 192.0.2.0/24 / 198.51.100.0/24 / 203.0.113.0/24 IPv4 (RFC 5737), 2001:db8::/32 IPv6 (RFC 3849), 00:00:5E:00:53:xx MAC (RFC 7042).

Argus Gateway integration: response headers should include X-Argus-Redact-Profile: pseudonym-llm; UI clients render display_text, LLM clients consume downstream_text. Storage of downstream_text as business truth is unsafe — it's synthetic by design.

Real users named like canonical fakes (e.g., a real customer named 张三 or John Doe): pass reserved_names={"person_zh": ()} (or person_en) to disable that locale's canonical-name pollution detection so the real user's name flows through normal redaction.

Streaming

For chat sessions or long-form input where text arrives in chunks, use StreamingRedactor (input side) and StreamingRestorer (output side). Both require complete logical units per chunk (sentence / paragraph / turn) — entities split across chunk boundaries are not handled.

from argus_redact.streaming import StreamingRedactor, StreamingRestorer

# Input side: redact each chunk; same original value across chunks → same fake
r = StreamingRedactor(salt=b"my-secret-salt", lang="zh")
for chunk in input_stream:                  # one sentence/paragraph/turn each
    res = r.feed(chunk)
    send_to_llm(res.downstream_text)

# Output side: restore LLM output stream at sentence boundaries
restorer = StreamingRestorer(r.aggregate_key())
for chunk in llm_output_stream:
    restored = restorer.feed(chunk)
    if restored:
        print(restored, end="")
print(restorer.flush(), end="")

True byte-level streaming (entities crossing chunk boundaries) needs full incremental detection and is roadmapped for a later release.

⚠️ Realistic-mode output must not be re-redacted (it would corrupt the key dict). redact_pseudonym_llm will raise PseudonymPollutionError if called on already-faked input — call restore() first.

Full API → · Design constraints →

Integrations

	Install
LangChain / LlamaIndex / FastAPI	core
Presidio bridge	pip install argus-redact[presidio]
MCP Server (Claude Desktop / Cursor)	pip install argus-redact[mcp]
HTTP API Server	pip install argus-redact[serve]
Structured data (JSON / CSV)	core
Streaming restore	core
Docker	slim 157MB / full 5GB

Security

PII never leaves your device. Per-message keys prevent cross-request profiling. Full security model →

Meets PIPL · GDPR · HIPAA technical requirements as a byproduct of its privacy-first design. Details →

Documentation

Getting Started	Install, first redact/restore, key management
API Reference	All parameters, return types, streaming, structured data
CLI Reference	Commands, flags, serve, MCP server
Configuration	Per-type strategies, enterprise mask rules, false positive reduction
Sensitive Info	Taxonomy, privacy levels, roadmap
PII Type Catalog	All PII types — strategy, sensitivity, PIPL/GDPR/HIPAA mapping (auto-generated)
Architecture	Three-layer engine, cross-layer hints, pure/impure separation
Language Packs	Adding new languages
Security Model	Threat model, compliance, per-message keys
PRvL Standard	Open evaluation standard: Privacy × Reversibility × Language
Layer 3 Benchmark	LLM model comparison, prompt design, regulatory analysis
Benchmarks	Evaluation against 9 public PII datasets
Performance	Latency, throughput, benchmark results

Contributing

CONTRIBUTING.md — language packs, test scenarios, framework integrations welcome.

Contributors

Who	Contribution
@aiedwardyi	Brazilian Portuguese language pack (CPF, CNPJ, phone)

License

Apache 2.0