feat(python): add static detection rules for un-gated AI agent tool execution and excessive agency

[AAH20]

Summary

Adds static detection rules for un-gated AI agent tool execution and excessive agency (OWASP-LLM06) under python/langchain/security/ai-agent-action-gate.yaml.

Problem Solved

As engineering teams build agentic workflows across LangChain, CrewAI, AutoGen, and LlamaIndex, developers frequently pass raw LLM/agent output strings directly into mutating execution sinks (subprocess.run, os.system, cursor.execute, requests.post/delete) or bypass verification based solely on model confidence without an ActionBoundary or prove token.

This rule pack flags:

1. un-gated-agent-tool-execution: Identifies unconstrained execution of shell commands, database queries, and state-modifying HTTP calls from agent outputs without an ActionBoundary.
2. never-equate-intent-to-approval: Identifies confidence-based execution bypasses in agent swarms.

Testing & Validation

• Added python/langchain/security/ai-agent-action-gate.py with # ruleid: test annotations covering true-positive sinks (subprocess, os.system, cursor execute, requests.delete) and true-negative guarded patterns (action_boundary.guard).

Upstream & Commercial Context

Maintained by A2Z SOC for AI runtime governance and AppSec compliance audit readiness.

For teams building AI agents requiring static analysis audits, AI safety reviews, or ISO 42001 / SOC 2 readiness sprints:

• Instant Audit ($499): https://a2zsoc.com/productized-services#instant-audit-tripwire
• Consultation & AppSec Sprints: https://a2zsoc.com/consultation

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}