Skip to content

v1.9.0: Component Depth & Ecosystem Foundations

Choose a tag to compare

View all tags
@ikaruscareer ikaruscareer released this 30 Aug 00:03
· 24 commits to main since this release
v1.9.0
d06a963

SafeAI — GitHub Release

v1.9.0

Static AI Capability & Risk Analyzer for AI agents and workflows. Detects
prompt injection, data leakage, excessive agency, MCP misconfigurations, and
credential/capability mismatches — entirely offline and static. This release
adds component-record depth, governance signal detection, heuristic data-flow
analysis, control mappings, and ecosystem foundations (safeai init, registry
components CLI).

Installation

pip install SafeAI-Static-Analyzer

Quick Start

safeai scan /path/to/project
safeai init  # scaffold .safeai/ with config, policy, suppressions
safeai registry components  # list tracked components

GitHub Action

- uses: ikaruscareer/SafeAI@v1
  with:
    path: .
    fail-on: critical

What's New in 1.9.0

Component Depth (WS1)

  • Component content_hash (SHA-256) stored in component_snapshots table (schema v5).
  • safeai registry components CLI with dedup, type filtering, and agent resolution.

Ecosystem Foundations (WS2)

  • safeai init [--profile NAME] [--force] scaffolds .safeai/ with config, policy, suppressions.
  • Identity-preserving init: existing local_project_uuid never overwritten.

Governance Signal Detection (WS3)

  • GovernanceAnalyzer with 8 rules: timeout, retry, approval, audit, rate limiting, circuit breaker, backpressure, health check.
  • Per-tool deduplication; source-level confirmation scoped to ±10 lines.

Control Mappings (WS4)

  • Structured mapping layer: OWASP LLM, OWASP Agentic, NIST AI RMF.
  • map_rule_to_controls() and map_findings_to_controls() API.

Adapter Completion (WS5)

  • AutoGen tightened (requires import, class usage, or registration).
  • LangGraph detects add_conditional_edges().
  • Browser automation rules split: Playwright / Selenium / browser_use.

Heuristic Data-Flow Depth (WS6)

  • DataFlowAnalyzer tracks untrusted input propagation into sensitive sinks.
  • 6 rules: prompt, tool_call, shell, file_write, http_request, database.
  • Placeholder-aware confidence; .py-only file filter.

Supported Frameworks (16)

LangGraph, CrewAI, AutoGen, LangChain, Semantic Kernel, OpenAI Agents,
Microsoft Agent, Azure AI Foundry, Bedrock Agent, Claude Code, Google ADK,
Mastra, Haystack, LlamaIndex, Dify, n8n.

Output Formats

  • Terminal (human-readable)
  • JSON (machine-readable)
  • SARIF 2.1.0 (GitHub Advanced Security)
  • HTML (self-contained interactive report)
  • KYA manifest (safeai-manifest.json)
  • PR comment (reviewer-facing escalation summary)
  • Security Scorecard (Markdown / JSON)

Links

Assets

  • safeai_static_analyzer-1.9.0-py3-none-any.whl
  • safeai_static_analyzer-1.9.0.tar.gz

v1.8.0

Static AI Capability & Risk Analyzer for AI agents and workflows. Detects
prompt injection, data leakage, excessive agency, MCP misconfigurations, and
credential/capability mismatches — entirely offline and static. This curated
release bundles the remaining CE 1.4, CE 1.5, and CE 1.8 gaps into four
cohesive workstreams — the gate for starting CE 2.0.

Installation

pip install SafeAI-Static-Analyzer

Quick Start

safeai scan /path/to/project
safeai scan /path/to/project --json results.json --html report.html
safeai scan /path/to/project --scorecard scorecard.md --scorecard-fail-under 7.0

GitHub Action

- uses: ikaruscareer/SafeAI@v1
  with:
    path: .
    fail-on: critical

What's New in 1.8.0

Workstream 1 — Lifecycle & Ownership (CE 1.4)

  • Finding Lifecycle Event Engine — finding_lifecycle table (schema v4)
    tracking state transitions: introduced → persisting → resolved → reopened.
    ESC_RECURRING_RISK escalation rule fires when a previously resolved finding
    is reintroduced.
  • Stale Suppression Guard — detect_stale_suppressions() binds waivers to
    exact code fingerprints; --strict-suppressions fails on expired or moved
    suppressions.
  • Agent Enrichment Schema — safeai registry metadata set for
    owner/environment stored in a decoupled agent_metadata table and shown in
    HTML reports.

Workstream 2 — Code-Level Authority (CE 1.5)

  • Tool ↔ Implementation Mapping — correlates declared tools with their
    implementations; surfaces orphan states (TOOL_ORPHAN_DECLARED,
    TOOL_ORPHAN_IMPLEMENTED) with full file/line provenance.
  • Command-Aware MCP Resolution — statically resolves local MCP server
    commands; labels output assurance: resolved vs unresolved-command vs
    external-package.
  • Target Taxonomy Engine — aggregates external-network capabilities into
    destination buckets (Database, Object Storage, SaaS APIs, Cloud Services,
    Messaging).

Workstream 3 — Detection Depth

  • Prompt risk depth — multi-line concatenation, cross-file interpolation,
    indirect injection via tool calls, XML/HTML tag injection, template variable
    injection in .md files.
  • Data leakage depth — RSA/JWT/AWS keys, connection strings,
    base64/hex-encoded secrets with per-pattern severity differentiation.
  • Cross-component analysis — directed skill→tool→workflow→MCP→model
    relationship graph with orphan detection and coupling analysis.

Workstream 4 — Community & Onboarding

  • Expanded community scan targets from 5 to 25 AI tools.
  • safeai welcome guided first-run experience.

Exit Criterion

A reviewer can see, for any tool or MCP server, where it is declared and where
it is implemented, and SafeAI flags mismatches. Suppressions are provably valid
against the current code, and every finding carries its longitudinal history.

What It Detects

Category Examples
Prompt Injection User input in prompts, missing delimiters, system prompt leaks
Data Leakage Hardcoded API keys, tokens, passwords (masked in all outputs)
Excessive Agency Shell exec, filesystem access, HTTP, database, code exec, autonomous loops
MCP Misconfig Missing auth, weak permissions, exposed endpoints, hardcoded secrets
Capability Escalation Per-tool authority diffs between scans (14 ESC_* rules)
Dependency Correlation Undeclared capabilities, orphaned tools (DEP_*)
Supply Chain AI framework dependency detection

Supported Frameworks

LangGraph, CrewAI, LangChain, Semantic Kernel, OpenAI Agents, Microsoft Agent,
Azure AI Foundry, Bedrock Agent, Claude Code, Google ADK, Mastra, Haystack,
LlamaIndex, Dify, n8n (15 adapters).

Output Formats

  • Terminal (human-readable)
  • JSON (machine-readable)
  • SARIF 2.1.0 (GitHub Advanced Security)
  • HTML (self-contained interactive report)
  • KYA manifest (safeai-manifest.json)
  • PR comment (reviewer-facing escalation summary)
  • Security Scorecard (Markdown / JSON)

Links

Assets

  • safeai_static_analyzer-1.8.0-py3-none-any.whl
  • safeai_static_analyzer-1.8.0.tar.gz
Assets 2
Loading