Skip to content

v2026.3.0 - Scanner Precision + FP Tuning

Choose a tag to compare

View all tags
@rosschurchill rosschurchill released this 16 Feb 20:41
· 94 commits to main since this release
v2026.3.0
5e306d2

What's New in v2026.3.0

Scanner Precision + FP Tuning — Compound scanner gates, attribution fix, and 508 FP filters.

Fixed

  • Scanner Attribution Bug — Parallel scan findings now correctly attributed to originating scanner (was all assigned to first scanner per file)
  • SteganographyScanner FPs — Compound gate: requires BOTH input handler AND AI keyword indicators
  • LLMGuardScanner FPs — Toxicity check now requires actual LLM API call patterns before firing
  • MCPServerScanner FPs — Tightened time-based, mcp-remote, and tool poisoning patterns
  • RAGSecurityScanner FPs — HTTP source narrowed to RAG-specific variable names
  • MultiAgentScanner FPs — Compound gate: requires framework imports/API calls (e.g. from crewai import, Agent()
  • Tool Poisoning Rule FPs — Tightened patterns in tool_attacks.yaml and mcp_vulnerabilities.yaml

Changed

  • 508 FP Filter Patterns (up from 430) — 96.8% false positive reduction on real-world projects
  • Compound Scanner Gates — SteganographyScanner, LLMGuardScanner, MultiAgentScanner use framework-aware gates

Removed

  • YAMLScanner (yamllint) — Style linter removed. YAML security covered by Trivy + Semgrep + MEDUSA built-in rules. Scanner types: 42 → 41.

Stats

  • 76 specialized analyzers
  • 3,000+ AI security detection patterns
  • 508 FP filter patterns (96.8% reduction)
  • 133 critical CVE detections
  • 38 files changed, 1,636 insertions, 178 deletions
Assets 3
Loading