memory-os

Commit Graph

Author	SHA1	Message	Date
ClaudioDrews	598eaf19d0	fix(sanitize): change system prefix replacement from silent removal to [REDACTED] - Silent removal (empty string) left grammatically broken sentences and no audit trail. [REDACTED] preserves context structure. - Control chars and zero-width Unicode keep empty string replacement (invisible characters, removal is correct). - Updated 4 test assertions in _test_sanitize.py.	2026-06-02 13:45:02 -03:00
David Soff	a5c1344afb	security(hooks, context_enhancer): add dual-layer prompt injection sanitization (#5 ) security(hooks, context_enhancer): add dual-layer prompt injection sanitization - Layer 1 (hooks.py): aggressive sanitization at prompt boundary with 11 regex patterns + heuristic. Detects override directives, template injection, URI schemes, system prefixes, HTML/XML, control chars, zero-width Unicode, and code fences. Applied to fabric, Qdrant, sessions, and facts injection points. Complements existing _is_system_injection() (which filters ingest, not egest). - Layer 2 (context_enhancer.py): lightweight sanitization on all search result content_preview outputs (hybrid, dense, sparse, lexical, sqlite). - _test_sanitize.py: manual validation script with 24 test cases. - Replacement strategy: [REDACTED] preserves audit trail and grammatical context instead of silent removal.	2026-06-02 13:43:34 -03:00

Author

SHA1

Message

Date

ClaudioDrews

598eaf19d0

fix(sanitize): change system prefix replacement from silent removal to [REDACTED]

- Silent removal (empty string) left grammatically broken sentences
  and no audit trail. [REDACTED] preserves context structure.
- Control chars and zero-width Unicode keep empty string replacement
  (invisible characters, removal is correct).
- Updated 4 test assertions in _test_sanitize.py.

2026-06-02 13:45:02 -03:00

David Soff

a5c1344afb

security(hooks, context_enhancer): add dual-layer prompt injection sanitization (#5 )

security(hooks, context_enhancer): add dual-layer prompt injection sanitization

- Layer 1 (hooks.py): aggressive sanitization at prompt boundary with
  11 regex patterns + heuristic. Detects override directives, template
  injection, URI schemes, system prefixes, HTML/XML, control chars,
  zero-width Unicode, and code fences. Applied to fabric, Qdrant,
  sessions, and facts injection points. Complements existing
  _is_system_injection() (which filters ingest, not egest).

- Layer 2 (context_enhancer.py): lightweight sanitization on all
  search result content_preview outputs (hybrid, dense, sparse,
  lexical, sqlite).

- _test_sanitize.py: manual validation script with 24 test cases.

- Replacement strategy: [REDACTED] preserves audit trail and
  grammatical context instead of silent removal.

2026-06-02 13:43:34 -03:00

2 Commits