Skip to content

SECURITY_AUDIT.md

Latest commit

 

History

History
95 lines (72 loc) · 5.06 KB

SECURITY_AUDIT.md

File metadata and controls

95 lines (72 loc) · 5.06 KB
Raw

Security Audit Checklist — AstroML / Fraud Registry

1. Smart Contract (Soroban / Rust)

1.1 Access Control

  • Admin-only functions (register_validator, update_config, deactivate_validator, update_validator_reputation) verify the caller matches the stored admin address
  • Non-admin callers receive Error::Unauthorized
  • FIXED (SC-1): initialize now has a guard against re-initialization using env.storage().instance().has(&DATA_KEY) check
  • Admin key rotation mechanism is not implemented; document the operational runbook for key compromise

1.2 Input Validation

  • confidence and reputation values > 100 are rejected with Error::InvalidInput
  • Boundary values 0 and 100 are accepted as valid
  • FIXED (SC-3): Empty reason string is now rejected with Error::InvalidInput
  • FIXED (SC-2): consensus_threshold of 0 is rejected with Error::InvalidInput in update_config

1.3 Replay / Duplicate Prevention

  • Duplicate reports from the same validator for the same account are blocked via Error::AlreadyReported
  • Unique validator counting in is_fraudulent prevents a single address inflating consensus

1.4 Sybil Resistance

  • Reputation minimum enforced before accepting reports
  • Configurable consensus_threshold requires independent validators
  • REVIEW: Admin can register unlimited validators and immediately set high reputations — document trusted-setup assumption or add a time-lock

1.5 Integer Safety

  • u8 arithmetic for reputation/confidence cannot overflow standard addition since values are validated to ≤ 100
  • u64 counters (report_count, accurate_reports) use saturating Soroban semantics
  • Confirm consensus_threshold comparison (validator_count >= data.consensus_threshold) uses matching integer types to avoid sign-extension issues

1.6 Storage

  • TTL / expiry of instance storage not configured — very old fraud reports persist indefinitely; consider archival strategy
  • Single DATA_KEY storage is atomic per ledger operation; no partial-write risk

1.7 Denial of Service

  • FIXED (SC-4): get_active_validators now accepts an optional limit parameter (default 100) to prevent unbounded iteration
  • get_fraud_reports iterates all reports per account — same concern for heavily-targeted accounts

2. Python ML Pipeline

2.1 Injection Attacks

  • All raw SQL queries must use parameterised statements (SQLAlchemy ORM or %s placeholders); audit astroml/db/ for string-formatted queries
  • Graph construction paths that accept external filenames must be validated against a whitelist of allowed directories

2.2 Secrets Management

  • config/database.yaml is listed in .gitignore (verify)
  • Ensure no credentials are hard-coded in source files (run git grep -n "password\|secret\|api_key")
  • Database passwords should be read from environment variables, not YAML files checked into VCS

2.3 Dependency Security

  • Run pip-audit against requirements.txt to identify known CVEs
  • Pin all dependency versions and maintain a lock file (pip-compile)
  • Rust dependencies: run cargo audit against Cargo.lock

2.4 Deserialization

  • Pickle-based model serialisation (torch.save / torch.load) must only load files from trusted paths; never load user-supplied model files directly

2.5 Data Leakage

  • Training labels must not be visible to the model during inference evaluation (covered by tests/test_leakage.py)
  • Logged metrics / artefacts must not contain PII from Stellar account addresses in plaintext

2.6 Configuration Security

  • Hydra / YAML configs must validate types and ranges on load; reject unknown keys
  • consensus_threshold and other thresholds in configs/ should have documented acceptable ranges

3. Infrastructure

3.1 Docker

  • Base images pinned to digest, not floating tags
  • Container does not run as root (USER directive set in Dockerfile)
  • No secrets in docker-compose.yml environment blocks in plaintext

3.2 CI/CD

  • Add cargo audit step to CI pipeline
  • Add pip-audit or safety check step to CI pipeline
  • Secret scanning (e.g., git-secrets or GitHub secret scanning) enabled on the repository

4. Remediation Tracker

ID Severity Finding Status
SC-1 High __init__ can be called again, overwriting admin Resolved
SC-2 Medium consensus_threshold = 0 marks all accounts fraudulent Resolved
SC-3 Low Empty reason string accepted Resolved
SC-4 Medium get_active_validators unbounded iteration Resolved
PY-1 High Confirm no hard-coded credentials in source Open
PY-2 High Run pip-audit; remediate CVE findings Open
PY-3 Medium Pickle load from untrusted path Open
IN-1 Medium Docker base image tags not pinned to digest Open