Skip to content

Graph-Native Infrastructure for Context and Accountable AI Systems

docs(changelog): add security-enhancement PR entries to [Unreleased]

KaifAhmad1ce60acb294d2
diff --git a/CHANGELOG.md b/CHANGELOG.md
index edf60092..a6b6c43f 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -7,6 +7,34 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
## [Unreleased]
+- **Security: 12 vulnerability fixes across CRITICAL → LOW severity** (PR `security-enhancement` by @KaifAhmad1):
+
+ **Critical**
+ - **Eval injection eliminated** (`semantica/parse/media_parser.py`): Replaced `eval(stream.get("r_frame_rate", ...))` — which executed arbitrary Python from ffprobe JSON output — with a `_safe_parse_fps()` helper using `fractions.Fraction`. No code execution possible regardless of ffprobe output content. (CWE-95)
+ - **Unsafe pickle deserialization replaced** (`semantica/context/agent_memory.py`): `AgentMemory.save()` / `load()` previously used `pickle.dump` / `pickle.load`, allowing RCE if an attacker could write the `.pkl` file. Replaced with `json.dump` / `json.load`. `MemoryItem.to_dict()` / `MemoryItem.from_dict()` added for safe round-trip serialization — `timestamp` via `isoformat()`, `embedding` dropped (not JSON-safe, regenerated on demand). Legacy `.pkl` files are detected and refused with a migration message. (CWE-502)
+
+ **High**
+ - **SQL injection hardened** (`semantica/ingest/snowflake_ingestor.py`): `WHERE`, `ORDER BY`, `LIMIT`, and `OFFSET` were f-string interpolated directly into Snowflake queries. `LIMIT` / `OFFSET` now use parameterized `%s` placeholders; `ORDER BY` is validated against a strict `^[A-Za-z_][A-Za-z0-9_]*(\s+(ASC|DESC))?` regex; `WHERE` clauses containing semicolons are rejected before execution. (CWE-89)
+ - **XXE protection added for RDF/XML parsing** (`semantica/explorer/utils/rdf_parser.py`): `rdflib.Graph().parse()` on XML-based RDF formats had no external-entity restrictions. Added `_safe_parse_rdf()` wrapper that calls `defusedxml.defuse_stdlib()` before parsing, neutralising Billion Laughs and local file-read XXE attacks. Falls back gracefully with a `UserWarning` if `defusedxml` is not installed. (CWE-611)
+ - **Security headers and CORS added to main server** (`semantica/server.py`): No CORS policy, no response security headers, and raw `Exception` details were returned to clients. Added `CORSMiddleware` (origins from `SEMANTICA_CORS_ORIGINS` env var, defaults to `localhost` only); `_SecurityHeadersMiddleware` emitting `X-Content-Type-Options`, `X-Frame-Options`, `X-XSS-Protection`, `Referrer-Policy`, `Permissions-Policy`, and HSTS (HTTPS only) on every response; global error handler that logs internally and returns a generic `500`. (CWE-346, CWE-200)
+ - **CORS and WebSocket hardened in Explorer app** (`semantica/explorer/app.py`): `allow_methods=["*"]` and `allow_headers=["*"]` narrowed to `GET, POST, DELETE, OPTIONS` and `Content-Type, Authorization` only. `KeyError` / `ValueError` exception handlers now log the real message server-side and return generic text to clients. WebSocket messages larger than 64 KB trigger close with code `1009` (Message Too Big), preventing memory exhaustion via large frame injection. (CWE-346, CWE-400)
+
+ **Medium**
+ - **Algorithm parameter validated by enum** (`semantica/explorer/routes/graph.py`): The `algorithm` query parameter previously accepted any string; unknown values silently fell back to BFS. Replaced with `_PathAlgorithm(str, Enum)` — FastAPI now returns `422 Unprocessable Entity` for any value other than `bfs` or `dijkstra`. (CWE-20)
+ - **RDF upload extension allowlist** (`semantica/explorer/routes/vocabulary.py`): No file extension check was performed before reading RDF uploads. Extension is now validated against `{".ttl", ".rdf", ".owl", ".xml", ".jsonld", ".json-ld", ".json"}` before any content is read. (CWE-434)
+ - **Prompt injection mitigated** (`semantica/semantic_extract/llm_extraction.py`): User-supplied text and entity/relation labels were embedded directly into LLM prompts via f-string interpolation — a crafted input like `"\n\nIgnore all above instructions..."` could override system instructions. All user-supplied content is now passed through `json.dumps()` before embedding, neutralising newlines, quotes, and instruction-override attempts. (CWE-1336)
+ - **Dynamic `__import__()` removed** (`semantica/pipeline/pipeline_validator.py`): `__import__("collections").Counter(...)` replaced with a proper `from collections import Counter` module-level import. (CWE-95)
+ - **ReDoS eliminated** (`semantica/explorer/routes/enrich.py`): `re.split(r"\s+AND\s+", antecedent_text, re.IGNORECASE)` on user-supplied rule strings could exhibit polynomial backtracking. Fixed by normalising whitespace first with `" ".join(text.split())` (no regex) then splitting on the literal `" AND "`. Closes CodeQL alert #12. (CWE-1333)
+ - **Path traversal blocked** (`semantica/server.py`): SPA catch-all route used `STATIC_DIR / full_path` without validation. Added `Path.resolve()` + `relative_to()` check that returns `400 Bad Request` for any path that escapes `STATIC_DIR`. Closes CodeQL alerts #13 and #14. (CWE-22)
+
+ **Low**
+ - **SPARQL result cap and timeout** (`semantica/explorer/routes/sparql.py`): SPARQL queries ran to completion with no row limit or timeout; expensive queries could exhaust memory or block indefinitely. Results are now capped at 5 000 rows; `asyncio.wait_for(..., timeout=30)` abandons the await after 30 seconds and returns a structured error response. A module-level `asyncio.Semaphore(4)` caps concurrent in-flight `graph.query` calls so that timed-out threads (which continue running in the pool) cannot crowd out other requests by exhausting executor workers. `SparqlResponse` gains a `truncated: bool` field so callers can detect a capped result set. (CWE-400)
+ - **Import upload size limit and extension allowlist** (`semantica/explorer/routes/export_import.py`): No file size or type checks were enforced before reading import uploads. Extension is now validated against `{".json", ".csv"}` (the formats the handler actually parses — allowlist trimmed to match implementation); a hard 50 MB cap is enforced before content is read. (CWE-434)
+
+ **CodeQL / scanning infrastructure**
+ - Added `.github/codeql/codeql-config.yml` with `paths-ignore` for `cookbook/**/*.html` and `cookbook/**/*.js`. Notebook-exported HTML files embed entire minified third-party bundles (Plotly + MapLibre GL JS v4.7.1) that triggered false-positive JS alerts #15–#18. The Advanced Setup workflow now references this config via `config-file:`. Closes CodeQL alerts #15, #16, #17, #18.
+ - Removed blanket rule-ID auto-dismiss job from `.github/workflows/codeql.yml`. The previous job dismissed every open alert whose `rule.id` matched a fixed list on each `main` push — this would silently suppress any future real vulnerability of the same type. Replaced with a commented template for pinning specific alert numbers when manual dismissal is genuinely required.
+
- **Fix: Knowledge Explorer — blockers and security hardening** (PR #420 by @ZohaibHassan16, review fixes by @KaifAhmad1):
- **Dockerfile**: Renamed `DockerFile` → `Dockerfile` (case-sensitive filename caused Docker build failures on Linux CI). Fixed `CMD` module path from the non-existent `semantica.server:app` to `semantica.explorer.app:app`, which caused the Docker image to crash on startup. Added `app = create_app()` at module level in `semantica/explorer/app.py` so uvicorn can reference the ASGI app instance directly.
- **CORS hardening**: Changed `EXPLORER_CORS_ORIGINS` default from `"*"` to `"http://localhost:5173,http://127.0.0.1:5173"`. Any deployment that does not explicitly set the env var no longer exposes the API to all origins. The env var override continues to work as before.