v4.0.0 — SecurityPlugin with Privacy Controls

What's New in v4.0.0

False-Positive Elimination

• Rewrote all SHAPE_RULES regexes — anchored to command position, eliminated bare keyword matches
• Removed 11 generic config filenames from sensitive file patterns (settings.py, config.json, etc.)
• Tightened DLP patterns (ICD-10, Passport, DEA) to require labeled context
• Workspace files now get path-only checks, no content DLP scanning
• Allow ssh-add, nc -z, localhost curl, grep in source code

Hook Architecture Upgrade

• Migrated from variable substitution to stdin JSON contract (--hook-event)
• Added hook audit logging (JSONL) for every hook activation
• LLM verification for low-confidence exec rules
• Quoted-context stripping to prevent false positives from pasted output

Security Controls (from securityagent-core)

• System prompt masking (hostname, OS, paths, username)
• Media attachment filtering (blocks image/audio/video before LLM)
• AES-256-GCM session encryption for audit data at rest
• Modifying output pipeline (redacts PII in-place vs blocking)
• Behavioral chain detection (credential exfil, bulk scan, privilege escalation)
• Session-based policy engine with rate limiting

Core Package Extraction

• endpoint_agent/, skills/, policy/ extracted to securityagent-core package
• SecurityAgent is now a thin OpenClaw integration wrapper
• Cloud CLI exfil patterns preserved (AWS S3, GCP Storage, Azure Blob)

Obsidian Memory Map (from v3)

• Cross-session memory via Obsidian vault (retained from v3.0.0)