safedata-guard 1.0.9

A lightweight framework for safely enabling LLMs to analyze pandas/Polars data without exposing raw data or blindly executing generated code.

safedata-guard

A lightweight framework for safely letting LLMs analyze pandas/Polars data without exposing raw rows or blindly running the code they generate.

Most "chat with your data" tools send the whole table to the model and run whatever code it writes, unchecked. safedata-guard fixes both halves: it sends a compact, quality-aware summary instead of raw rows, and runs the model's code behind guardrails on a copy of your data.

Status: beta. Useful and tested, but treat it as a defense-in-depth safety layer, not a hardened sandbox. It is not a "fully secure sandbox", "compliance-grade PII protection", or "guaranteed safe execution" - PII detection and code screening are best-effort heuristics (see Scope below). For untrusted code, run it inside OS-level isolation (isolation="docker" or your own container/VM).

The recommended path

Pick a policy for your data and let safe_answer do the rest. It builds the minimum safe view for the question, runs the model's code behind the guardrails, and returns the answer plus an audit:

import safedata as sd

policy = sd.Policy.regulated()   # PII firewall + redaction + k-anonymity + caps

out = sd.safe_answer(df, "What is total revenue by region?",
                     model=my_llm, policy=policy)
print(out["answer"], out["audit"])

Profiles: Policy.basic(), Policy.regulated() (customer/PII data), Policy.strict() (container isolation + Presidio), Policy.audit_only(). Any field can be overridden: Policy.regulated(min_group_size=10).

Policy.strict() needs Docker (isolation="docker", a prebuilt runner image) and the optional Presidio install (use_presidio=True). Both degrade gracefully if absent - Presidio is skipped, and Docker raises a clear error - but if you want strong defaults without Docker, use Policy.regulated() (process isolation, k-anonymity, deep PII scan).

safe_answer + Policy is the recommended entry point. The pieces below (Agent, run_safely, create_contract, privacy_report, ...) are the lower-level building blocks it is composed from, for when you need finer control.

What it does

1. Summarises before the data reaches the model. Instead of 100,000 rows, it sends columns, types, a few sample values, basic stats, and warnings about common data traps: numbers stored as text ("$500"), the same category written several ways ("North"/"north "), dates-as-text and Excel serial dates (45292), non-unique IDs, empty/mostly-empty/constant columns, duplicate column names, and unexpected negatives.

2. Runs the model's code behind an AST screen. Before running, a static screen refuses anything outside in-memory analysis:

• imports beyond a small set (pandas, numpy, math, statistics, datetime, re)
• introspection/dunder tricks and dangerous builtins
• file/data readers and writers, however reached: read_*/to_*/write_* methods, file-backed classes (ExcelFile, ExcelWriter, HDFStore), aliases (w = df.to_csv), direct imports (from numpy import save), SQL readers, and internal helpers behind pd.io.* / np.lib.* / np.ctypeslib / np.f2py
• the df.eval() / df.query() string channels the screen can't inspect

It then runs on a copy of your data in a separate process with a timeout. The model may add/transform columns freely; afterwards the guardrail checks it didn't silently drop rows (unless allow_row_reduction=True) or return an empty result, and feeds any error back so the model fixes its own code.

Scope: please read honestly

This is defense in depth for cooperative or semi-trusted model output: it stops the destructive accidents an honest model makes and the obvious escape attempts. It is not a sandbox for deliberately malicious code. In-process Python screening can be defeated, and a child process still shares your filesystem permissions, so isolation here means timeout + crash safety, not a filesystem jail. For untrusted code, run inside OS-level isolation (container, locked-down user, or VM). PII masking and quality checks are best-effort heuristics, not a compliance guarantee.

Hardened isolation for untrusted code

The default (isolate=True) runs in a separate process with a timeout; crash and hang safety, but the child still shares your filesystem permissions. For genuinely untrusted model output, switch to container isolation:

Build the runner image once (it bundles safedata + pandas/numpy so the container needs no network at run time; see the repo Dockerfile):

docker build -t safedata-guard-runner:1.0.9 .

agent = safedata.Agent(model=..., isolation="docker",
                       memory="512m", cpus="1.0", network="none")
# or directly:
safedata.run_safely(code, df, isolation="docker")

The container runs with no network, a read-only root filesystem, and memory/CPU caps; only a throwaway work directory is writable. The image must already contain safedata (the locked-down defaults make a run-time pip install impossible by design); point at your own with docker_image=.

Guarding the result

Stop generated code from handing back the entire table (or raw sensitive rows):

safedata.run_safely(code, df,
                    max_result_rows=50,        # block oversized results
                    max_result_bytes=1_000_000,
                    redact_result_pii=True)     # scrub PII from the answer

Oversized results are blocked with a message telling the model to aggregate, rather than silently truncated. The same options are accepted by Agent(...).

Limitation: redact_result_pii works on DataFrames/Series (it knows the column) and on emails/phones in strings, but once names are flattened into a plain list (df['customer_name'].tolist()) the column context is lost and regex can't tell a name from any other text - so names can still leak that way. The robust defence is the column firewall (blocked_columns= / Agent.safe()), which masks unneeded PII columns before the code runs, so the values aren't there to leak in the first place.

Secure presets

The secure configuration is one call away, so you don't have to remember the flags:

agent = safedata.Agent.safe(model)     # result caps + PII redaction, process isolation
agent = safedata.Agent.strict(model)   # same, but runs code in a locked-down container

Any keyword overrides the preset (e.g. Agent.safe(model, timeout=30)).

Data Safety Contract (lower-level)

Most users want safe_answer + Policy (above), or create_privacy_plan() (the modern firewall plan). create_contract() is the lower-level building block it composes; reach for it when you want the raw allowed/blocked policy without building a safe view.

Turn the read-only checks into a machine-readable policy you can gate AI access on (no code is run). Pass the question to get a least-privilege firewall - the PII columns the question doesn't need are blocked:

contract = safedata.create_contract(df, question="total revenue by region")
# {"allowed_columns": ["revenue","region",...], "blocked_columns": ["email","customer_name"],
#  "data_traps": [...], "max_result_rows": 50, "privacy_level": "strict", ...}

safedata.run_safely(code, df, blocked_columns=contract["blocked_columns"])
# refuses code that touches a blocked column:
#   Blocked: the code accessed restricted column(s) the question does not need: email

Agent.safe() / Agent.strict() enable this firewall automatically. Add enforce_minimal_result=True to also refuse a full-table answer to an aggregate question.

Query-aware privacy firewall (minimum safe data)

Instead of asking "is this DataFrame safe?", ask "what is the minimum safe data needed to answer this question?" The firewall builds a privacy-filtered view, runs the analysis only on it, and returns an audit of what was dropped.

plan = safedata.create_privacy_plan(df, "total revenue by region")
safe_df = safedata.make_safe_view(df, plan)        # privacy-filtered safe view
answer  = safedata.safe_answer(df, "total revenue by region", model=my_llm)
print(answer["answer"], answer["audit"])

Safe by default. safe_mode="drop_unneeded_pii" (the default) removes the PII columns the question doesn't need and keeps every non-PII column - so the model never sees unneeded names/emails, but the columns your analysis needs are always present (no wrong answers). The audit explains exactly what was dropped:

Dropped 2 unneeded PII column(s) (customer_name, email) - never sent to the model. Retained 4 column(s) for analysis: ...

safe_mode="minimal" is an opt-in advanced mode that also drops non-PII columns the question doesn't appear to reference. It's stronger, but column relevance is a heuristic - if it misses a column your analysis needed, the answer can be wrong, so it carries a warning and is never the default.

k-anonymity (stop singling-out)

A group of one re-identifies a person ("average salary by postcode" where a postcode has one resident). Set min_group_size and grouped results must carry a per-group count; groups below the threshold are suppressed - rows removed, the figures that remain are exact (no number is fudged):

safedata.run_safely(code, df, min_group_size=5)
safedata.safe_answer(df, "average salary by postcode", model=my_llm, min_group_size=5)
agent = safedata.Agent.safe(model, min_group_size=5)
# or the standalone utility on any grouped frame with a count column:
safedata.k_anonymize(grouped_df, min_group_size=5)

If a grouped result has no count column, it's refused with guidance (the model adds counts and retries). (Private surrogate filters - masking an individual lookup behind a match flag - remain deferred until they can be done properly.)

From the CLI, preview the safe view for a question without running anything:

safedata plan customers.csv "total revenue by region" --json

The plan reports dropped_pii_columns, allowed_columns, the audit, and both original_risk_level and safe_view_risk_level (the raw frame can read "high" while the exposed view is "low" because PII was dropped first).

Is it safe to send this to an AI?

safedata.ai_risk_score(df, "total revenue by region")
# {"risk_level": "high", "score": 65, "recommended_mode": "strict",
#  "reasons": ["High-sensitivity PII columns: email", ...]}

safedata.detect_ai_traps(df)   # traps that make an AI answer wrong, with fixes
safedata.shadow(df)            # synthetic same-shape frame, no real values

On the CLI: safedata risk customers.csv "What is total revenue by region?" (exit code 2 on high risk, so it can gate a pipeline).

Audit trail for an answer

Every agent.ask() result can write a self-contained HTML audit - the question, the exact summary sent to the model, each attempt (and why any were blocked), the final code/answer, data-quality warnings, withheld PII columns, and token saving:

out = agent.ask(df, "What were total sales in 2025?")
out.audit_report("audit.html")

Install

pip install safedata-guard
pip install "safedata-guard[polars]"   # optional, for Polars support

Core APIs (summarize, run_safely, Agent, validate, tokens) support pandas and Polars; the library detects the type. The HTML report() currently supports pandas (pass a Polars frame through df.to_pandas() first).

Quick start

import safedata, pandas as pd

df = pd.DataFrame({"date": ["2025-01-01", "2024-05-01", "2025-08-01"],
                   "amount": [100.0, 50.0, 200.0]})

def my_model(prompt):          # plug in any LLM: text in, code out
    return "result = df[df['date'].str.startswith('2025')]['amount'].sum()"

agent = safedata.Agent(model=my_model)
out = agent.ask(df, "What were total sales in 2025?")
print(out.answer)              # 300.0
print(out.blocked, out.attempts, out.tokens)

Connecting a real model

Real models return messy text (Markdown fences, chatter, occasional failures). safedata.wrap() takes any text-in/text-out function, extracts the bare code, and raises a clear ModelError on failure, so you're not tied to one provider.

def my_call(prompt):
    return some_model_that_takes_and_returns_text(prompt)   # OpenAI, local, ...

agent = safedata.Agent(model=safedata.wrap(my_call))
out = agent.ask(df, "What were total sales in 2025?")

A stronger model just means good code on the first try and fewer retries; the safety guarantees do not depend on it.

Token saving

Sending a whole table costs tokens per row; the summary is far smaller. As a rough illustration, a 1,000-row table estimates at ~18,180 → ~229 tokens (~98.7%) for one question; on millions of rows the saving approaches 99.99%.

print(safedata.token_savings(df))    # readable sentence
safedata.token_stats(df)             # {summary_tokens, raw_tokens, saved_*}

These are estimates, not tokenizer-exact counts: the library uses a provider-agnostic ~4-characters-per-token heuristic and sizes the raw data from a small row sample (it never serialises the whole table, so it stays cheap on huge frames). Exact numbers vary by model/tokenizer - treat the figures as orders of magnitude, not guarantees.

Plug into Pandera / Great Expectations

safedata is the AI-safety layer; it connects to the schema and quality tools your team already uses instead of replacing them. Export the inferred schema and checks:

schema = safedata.to_pandera_schema(df)            # a pandera DataFrameSchema
schema.validate(df)

suite = safedata.to_great_expectations_suite(df)   # a portable GX suite dict
# {"expectation_suite_name": ..., "expectations": [...],
#  "meta": {"pii_columns": ["customer_name", ...]}}

to_pandera_schema needs pip install "safedata-guard[pandera]". to_great_expectations_suite returns a plain dict that Great Expectations can import, so safedata doesn't pin you to a specific (fast-moving) GX version.

International PII (optional, Presidio)

The built-in detector (regex + name hints + Luhn) is strong for US/UK-style data but misses names in free text, locations/addresses, IBANs and non-US formats. For international coverage, install the optional extra and turn it on:

pip install "safedata-guard[presidio]"
python -m spacy download en_core_web_sm

safedata.enable_presidio()        # call once; whole pipeline now uses it
safedata.privacy_report(df)       # also flags names/locations/IBANs, etc.

It augments (never replaces) the built-in detector and feeds the firewall, contract, risk score, and Agent. Off by default, so there's no heavy dependency unless you opt in.

PII masking

The summary includes a few real sample values, which can contain personal data. By default safedata masks obvious PII (emails, cards, phones, SSNs, IPs) before the summary leaves your machine and notes which columns were masked.

safedata.summarize(df)                    # regex PII (emails/cards/...) masked
safedata.summarize(df, mask_pii=True)     # ALSO withhold name/address columns
safedata.summarize(df, redact_pii=False)  # raw samples, if you are sure

Note: plain summarize(df) masks only regex-detectable PII (emails, cards, phones, SSNs, IPs) - it does not hide names/addresses, so its raw output can still contain "Alice Smith". Pass mask_pii=True (or use build_safe_prompt() / Agent.ask(), which do this for you) before sending a summary to a model.

Regex masking cannot catch names or addresses; build_safe_prompt(..., privacy= "mask") (below) goes further and fully withholds every detected PII column.

Agent.ask() does this withholding by default (mask_prompt_pii=True): name and address columns are masked in the summary the model sees and in the audit report, not just regex-matchable emails. Column names/types are still shown, so the model can still operate on those columns. With redact_result_pii=True, the returned value is also scrubbed - PII columns of a result frame are replaced with [REDACTED], and dict/list results are walked recursively.

Data quality & AI-readiness API

The same findings are also available as structured objects you can act on, each with a rule id, severity, confidence, column, evidence, and (where possible) ready-to-run fix code.

import safedata as sd

sd.validate(df)          # list[Issue]: rule_id, severity, confidence, evidence...
sd.suggest_fixes(df)     # [{issue, column, suggested_code}], runnable pandas
sd.explain_issue(issue)  # plain-language explanation
sd.quality_score(df)     # {score 0..100, breakdown, privacy_risk}
sd.ai_readiness(df)      # {ready_for_summary, safe_to_send_raw, needs_review, ...}
sd.privacy_report(df)    # {pii_columns, high_risk, medium_risk, actions}
sd.infer_columns(df)     # {col: "identifier"|"date"|"money"|"pii_email"|...}
sd.build_safe_prompt(df, "What are the top trends?", privacy="mask")

validate() is read-only and never runs code. quality_score().privacy_risk is driven by the kind of PII found (one email column = High), kept separate from the data-quality number. build_safe_prompt(privacy="mask") withholds all PII columns, including the name/address columns regex cannot see, so they never reach the model.

Command line

safedata check sales.csv                     # summary + quality score + tokens
safedata check data.xlsx --report out.html   # also write an HTML report
safedata check sales.csv --no-redact --samples 5
safedata check sales.csv --json              # machine-readable for automation
safedata check customer.csv --fail-on pii    # exit 2 if PII present
safedata check sales.csv --fail-on high      # exit 2 on any high-severity issue

--json emits quality_score, privacy_report, ai_readiness, issues, pii_columns, tokens. --fail-on (low/medium/high/pii/any) turns safedata into a gate for CI/CD, Airflow, or pre-refresh checks. The CLI only reads and summarises; it never executes model code. Supported formats: .csv, .tsv, .xlsx, .xls, .parquet, .json. Also runs as python -m safedata check ....

Function reference

Agent loop

• Agent(model, max_retries=3, isolate=True, timeout=10.0, allow_row_reduction=False) (isolate/timeout/allow_row_reduction pass through to run_safely).
• agent.ask(df, question, verbose=False) → result with .answer, .blocked, .reason, .attempts, .tokens.

Connecting a model: wrap(call, clean=...), extract_code(text), ModelError.

Running code safely

• run_safely(code, df, result_var="result", isolate=True, isolation=None, timeout=10.0, allow_row_reduction=False, max_result_rows=None, max_result_bytes=None, redact_result_pii=False, **docker_opts) runs on a copy, blocks unsafe ops, checks invariants and result-size/PII guards, returns the result. Raises SafetyError. isolation="docker" runs in a locked-down container; if the subprocess runner is unavailable, the in-process fallback still enforces timeout via a thread.
• check_code(code) → CodeCheck(.safe, .reason); screens without running.

Looking at the data: summarize(df, redact_pii=True, mask_columns=None), report(df, path=None).

Structured analysis: validate, Issue, suggest_fixes, explain_issue, quality_score, ai_readiness, privacy_report, infer_columns, build_safe_prompt.

Tokens: token_savings(df), token_stats(df), estimate_tokens(text).

Development

Run the test suite with the dev extras installed (they include pytest-timeout, polars, and openpyxl so the full suite and its config apply):

pip install -e ".[dev]"
pytest -q

Running a bare pytest without the dev extras still works, but prints a harmless Unknown config option: timeout warning because the optional pytest-timeout plugin isn't present.

License

MIT