argus-redact 0.6.8

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.

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.

56 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.
• 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) — M1 Max, mode="fast":

Text	redact()	restore()	Throughput
Short (17 chars)	0.07ms	0.04ms	13,036 docs/sec
Medium (770 chars)	1.00ms	0.05ms	1,031 docs/sec
Long (10K chars)	22.2ms	0.05ms	45 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)	78.3%	30.3%	43.7%
ner (+ spaCy)	72.8%	41.4%	52.8%
auto (+ Ollama 32B)	skipped this run

ai4privacy en, 500 samples, v0.6.6. 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 (v0.6.8)	Next milestone
Protected	56 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)	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