OWASP Top 10 for Agentic Applications (Dec 2025) compliance enforcement. Wraps builtin_tools/compliance.py — prompt-injection detection/blocking and excessive-agency limits (max tool calls + task duration).

How it works

Two enforcement modes:

Prompt injection detection

Scans user prompts for injection patterns (embedded instructions, role-playing attacks, context injection). Matches are blocked or flagged based on workspace config.

Excessive agency limits

Configurable caps on:

Maximum tool calls per task (max_tool_calls)
Maximum task duration in seconds (max_task_duration_s)

Breaches trigger a pause and human review request.

Install

In org template (org.yaml)

plugins:
  - molecule-compliance

From URL (community install)

github://Molecule-AI/molecule-ai-plugin-molecule-compliance

Configuration

In workspace config.yaml:

compliance:
  injection_mode: block   # or: warn
  max_tool_calls: 50
  max_task_duration_s: 600

Runtimes

langgraph — primary
claude_code — supported
deepagents — supported

Skills

owasp-agentic — agent-side guidance on OWASP Top 10 risks and mitigation patterns

Architecture

skills/
  owasp-agentic/
    SKILL.md   # Agent-side guidance on OWASP compliance gates
adapters/
  claude_code.py  # Installs skill via AgentskillsAdaptor; compliance hooks run as
                  # pre-call gate inside the skill loader, not agent-callable tools

Known issues

See known-issues.md.

README.md

molecule-compliance