Excessive Data Exposure (Python)

ID

excessive_data_exposure_python

Severity

high

Remediation Complexity

medium

Remediation Risk

high

Remediation Effort

medium

Family

API3:2023 - Broken Object Property Level Authorization

CWE

CWE-213, CWE-200

Resource

data_exposure

Language

FastAPI, Flask, Django / DRF

Description

Walks each endpoint’s response model and classifies every field into one of three confidence tiers, based on whether the field is sensitivity-tagged (PII / PCI / PHI / credentials by the sensitivity classifier) and whether the same field name appears in the request:

  • HIGH — sensitive AND not referenced in the request. The caller never asked for the field; returning it is over-fetch.

  • MEDIUM — sensitive AND referenced in the request. Possibly a legitimate field-by-field update — surfaced for review.

  • LOW — not sensitivity-tagged AND not referenced in the request. Mild signal of over-fetching.

Only the HIGH tier fires by default. The detector is per-language because the framework idioms differ — JSON-annotation models in Java / C#, dataclass / Pydantic models in Python, plain-object / class-transformer in JS/TS, struct tags in Go, Eloquent / Symfony serializer groups in PHP.

Rationale

Excessive Data Exposure is the read side of OWASP API3:2023 — the API returns more than the client needs, because the implementation returned a persistence model directly and the persistence model has more fields than the contract. The risk is twofold:

  • Direct data leak — the extra fields contain credentials (passwordHash, apiToken), PII (email, phone, ssn), or PCI (cardLastFour, accountNumber). Anyone reading the response sees them.

  • Object-property authorization bypass — even if the endpoint is authorised to return the object, individual fields on that object should be scoped. A "GET my profile" endpoint authorised for the user should not expose the user’s isAdmin flag, lastLoginIp, or internalNotes. API3 separates these from API1 (object-level) for exactly this reason.

The pattern is endemic in single-page-app backends, where the same User / Order / Patient entity is reused across all read endpoints, with the client deciding which fields to show. The "client filters out the bad fields" defence does not work — anyone can read the network response.

A susceptible FastAPI handler might look like this:

class User(SQLModel, table=True):
    id: int
    username: str
    email: str            # PII
    password_hash: str    # CREDENTIAL
    ssn: str              # PII
    is_admin: bool        # PRIVILEGE

@app.get("/users/{user_id}")
def get_user(user_id: int) -> User:
    return session.get(User, user_id)

Returning the ORM model directly serialises every field — including password_hash, ssn, and is_admin.

Remediation

The fix is to return a dedicated response model per operation, narrow to the fields the operation actually needs to expose. Specific per-language patterns are in the language pages.

Beyond per-route DTOs:

  • If the sensitivity tag is wrong for your project — for example, an email-newsletter app whose User.email field is intentionally public — exclude the field via per-project sensitivity overrides rather than suppressing the finding.

  • Pair this detector with pii_leak_in_response (focuses on PII / PCI / PHI specifically, with stricter severity).

  • Server-side response filtering is the only effective control. Anything that depends on the client hiding fields is not a remediation.

Here is a revised handler with an explicit response model:

class PublicUser(BaseModel):
    id: int
    username: str
    model_config = ConfigDict(extra='forbid')

@app.get("/users/{user_id}", response_model=PublicUser)
def get_user(user_id: int):
    return session.get(User, user_id)   # FastAPI filters to PublicUser fields

response_model=PublicUser tells FastAPI to project the result onto the declared model — fields not in PublicUser never reach the wire, even if the ORM model gains new sensitive columns.

DRF equivalent: a serializer with an explicit fields = ('id', 'username') (never all).

Flask equivalent: build the response dict manually rather than calling jsonify(model.dict).

Configuration

The detector accepts:

  • minConfidence — the minimum confidence tier that fires. Default high. Set to medium to include "sensitive AND referenced in request" findings. Set to low only for benchmark / audit runs (very noisy).

The set of sensitivity tags (PII / PCI / PHI / credentials) is configured globally on the sensitivity classifier, not per-detector.

References