v0.2.2 — CLI version string fix

What's fixed

• prism --version now correctly reports Prism Scanner v0.2.2. The v0.2.1 release shipped with cli.py still hardcoding VERSION = "0.1.3", causing a confusing version mismatch after pip install -U prism-scanner.

What's NOT changed

This is a cosmetic-only patch. Detection rules, engine behavior, and performance are 100% identical to v0.2.1. If you've already upgraded to v0.2.1 and don't care about the version banner, you don't need to upgrade.

Recap: what's new in the v0.2.x line (since v0.1.3)

If you're coming from v0.1.x, here's what you get:

• P10 — Agent psychological manipulation detection: gaslighting, guilt-tripping, authority impersonation, urgency pressure, emotional coercion. Based on Northeastern/Harvard/MIT research; to our knowledge the first open-source automated detector for this attack class.
• M7 — Publish hygiene checks: source maps, .env files, private keys, IDE configs, package-manager credentials. Inspired by Anthropic's April 2026 Claude Code source map leak. Wire it into your CI:
  ```bash
  prism scan ./dist --fail-on high
  ```
• 47–93% fewer false positives on real-world skills (S1, S4, S8, S12 engine fixes; validated against 3 representative ClawHub skills).
• 41 detection rules total across S1–S14 (behavior), M1–M7 (metadata + publish hygiene), P1–P10 (patterns + manipulation), R1–R10 (residue).

Install

```bash
pip install -U prism-scanner
prism --version # → Prism Scanner v0.2.2
```

Full changelog

v0.2.1...v0.2.2

v0.2.1: False Positive Reduction

What's Changed

Systematic false positive reduction validated against real ClawHub skills. This is a quality-only release — no new detection rules, only existing rules made more accurate.

Key Improvements

Fix	Before	After
S1 uvicorn.run()/asyncio.run() misfire	Flagged as shell exec	Requires subprocess./os. prefix
S12 path.open("rb") misfire	Flagged as deserialization	Only fires on pickle./yaml./marshal./shelve.
P5 MAL-010/011/012 cross-file matching	Whole-file multiline → CRITICAL	Same-line proximity matching
P3 regex escapes / magic bytes	[\x00-\x1f] in regex flagged	Regex context & binary headers excluded
P4 version numbers / doc IPs	4.1.6.14 flagged as IP	Version context & RFC 5737 IPs filtered
S5 config env vars noise	59 LOW findings for config vars	Non-sensitive → INFO (no grade impact)
M7 core/ directory	Python package dir flagged as core dump	Directories skipped

Validation on Real Skills

Skill	v0.1.3	v0.2.1	Reduction	Grade
pitertxus/pensieve (blockchain)	211 findings	18	-91%	F → D
haiyangchenbj/invassistant (finance)	179 findings	12	-93%	F → D
larryfang/em-intel (engineering)	118 findings	62	-47%	F → F*

*em-intel remains F due to a genuine data exfiltration chain correctly identified by the scanner.

Testing

97 tests, all passing. 9 new regression tests for FP/TP coverage.

Full Changelog: https://github.com/aidongise-cell/prism-scanner/blob/main/CHANGELOG.md

v0.1.0 — Initial Release

🔷 Prism Scanner v0.1.0

Security scanner for AI Agent skills, plugins, and MCP servers.

Highlights

• 39 detection rules across 3 analysis layers:
  • Code Behavior (S1-S14): Shell execution, data exfiltration, SSRF, persistence, unsafe deserialization, download-and-execute
  • Metadata (M1-M6, P1-P9): Hardcoded credentials, typo-squatting, obfuscated payloads, prompt injection
  • System Residue (R1-R10): LaunchAgents, crontab, shell config pollution, credential leaks
• Lightweight taint analysis — tracks source-to-sink data flows within each file
• Letter-grade risk assessment (A-F) with transparent scoring logic
• 4 output formats: Terminal (rich), JSON, HTML, SARIF (GitHub Code Scanning)
• Remote scanning via git clone with depth/size limits
• Multi-platform: ClawHub, MCP, npm, pip
• Residue cleanup: scan → plan → apply → rollback
• CI/CD ready: --fail-on severity gating + GitHub Action
• MCP Server mode: expose scanning tools to AI assistants
• Offline mode: --offline for air-gapped environments
• 67 automated tests covering all engines, scoring, and CLI

Install

pip install prism-scanner

Links

• PyPI
• Documentation
• ClawHub Top-100 Security Report