Skip to content

Security: abk1969/ai-act-skills

Security

SECURITY.md

Security policy

Resource profile

This repository ships a decision-support agent skill. The skill declares (in skills/ai-act-compliance/ssl.json):

Property Value
permission filesystem.read (own reference files only)
touches_sensitive_resources false
control_flow_features.tool_calls false
Network access none
Credentials access none
Code execution none
External tool calls none

Verification:

# Manifest declares no tool calls
jq '.scheduling.control_flow_features' skills/ai-act-compliance/ssl.json

# Skill body confirms — zero tool invocations
grep -nE "Bash\(|Edit\(|Write\(|Read\(|WebFetch\(" \
  skills/ai-act-compliance/SKILL.md \
  skills/ai-act-compliance/references/*.md
# → no matches

The skill's only side effect is reading its own bundled reference files to inform its textual output. It cannot write, network, exec, or call host tools. This makes it a low-risk addition to any agent runtime.

Supported versions

We provide security updates for the latest minor release.

Version Supported
1.2.x ✅
1.1.x ⚠️ critical-only
1.0.x ❌

Reporting a vulnerability

If you find a security issue, please do not open a public issue. Use one of:

  1. GitHub private vulnerability reporting
  2. Email the maintainer (see commit history for the address)

We will:

  • Acknowledge receipt within 72 hours
  • Triage and confirm the issue within 7 days
  • Ship a fix within 30 days for high/critical issues; 90 days for medium

Threat model

In scope for this repo's security policy:

  • Prompt injection attacks via crafted description or reference-file content that could mislead a host LLM into bypassing the skill's scope (e.g., emitting non-AI-Act content, ignoring the legal disclaimer, leaking system prompts)
  • Supply-chain attacks: tampered ssl.json, malicious commits to reference files, typosquatted forks impersonating this repo
  • Citation-tampering: deliberately wrong article numbers, ISO clause references, or sanction figures that could lead a deployer to misclassify their AI system
  • Discovery-file abuse: malicious AGENTS.md / GEMINI.md content that misroutes a host runtime

Out of scope for this repo:

  • Vulnerabilities in the host agent runtime (Claude Code, Gemini CLI, OpenAI Codex) — report to the respective vendors
  • Vulnerabilities in third-party tools the user combines this skill with (MCP servers, benchmark frameworks, document generators)
  • Issues with the user's own AI system — this skill is decision-support for their compliance work, not a runtime safeguard
  • Misinterpretation of the AI Act when the cited articles and clauses are accurate (consult qualified counsel — that's expressly outside the scope of decision-support)

Snyk × Vercel agent-skill ecosystem

This skill is published to skills.sh and benefits from the Snyk × Vercel agent-skill ecosystem security work. The skill's minimal resource profile (no network, no credentials, no tool calls) is designed to clear those reviews trivially.

Related

There aren't any published security advisories