English | Русский

Security updates are applied to the latest minor version:

Verification semantics changed incompatibly in 0.9.1: documents with a fabricated, uncorroborated requisite are now reported as discrepancy_detected instead of zero_trust_verified, and the partially_verified and gate_not_executed statuses were introduced. Consumers checking for zero_trust_verified alone must review the new statuses before upgrading.

ScanReader is designed to process legally sensitive court documents, enforcement orders, and personal data (PDn). The architecture enforces the following security boundaries:

• ScanReader is capable of operating completely offline in a closed network perimeter using local VLM engines (e.g. Ollama or vLLM running Qwen2.5-VL).
• External network requests to OpenAI or cloud providers occur only when explicitly configured via environment variables (OPENAI_BASE_URL).

• Checksums and validations are calculated deterministically in-process without transmitting data to third parties.
• Logging utilities employ automatic regex-based masking (mask_secret) for sensitive strings, API keys (sk-..., Bearer tokens), and Russian passport numbers.

• All input files and MCP payloads are bounded by MAX_INPUT_BYTES = 50 MB to prevent memory exhaustion, out-of-memory crashes, and denial-of-service vulnerabilities.
• Token Bucket rate limiters protect local GPU instances from concurrency-induced thrashing.

• File writes employ write_atomic with temporary files, explicit disk syncing (os.fsync), and atomic replacement (os.replace) to guard against corruption during power interruptions.

If you discover a security vulnerability within ScanReader, please do not file a public issue. Instead, report it directly to the security maintainer:

• Telegram: @riper21
• Response Time: Reports are acknowledged within 48 hours, remediation timelines are provided within 5 business days.