EC-Council’s ADG 2.0 Pushes AI Governance From Policy to Proof

Most companies deploying AI may have a governance policy on file but ask the security leaders what an AI agent did in production last week, and under whose authority, and the answer usually can get messily vague.

EC-Council is aiming at that gap with ADG 2.0, the second version of its Adopt, Defend, Govern framework. It’s a big update. The original 12 minimum controls are gone, replaced by more than 180 controls in 12 control families. Those controls map to more than 90 regulations, standards and frameworks, including the EU AI Act, NIST AI RMF, ISO/IEC 42001 and MITRE ATLAS.

AI agents now call on internal tools and act on company data with little human input, which raises questions for security teams. Who watches an agent once it’s running? If it oversteps, who can shut it down?

Those questions stopped being hypothetical in July. OpenAI disclosed that models undergoing an internal cybersecurity evaluation had circumvented their testing environment and compromised parts of Hugging Face’s production infrastructure. Governments, meanwhile, are split on how to regulate AI; no global treaty is in sight.

“Most AI governance today is a policy document nobody tests,” said Jay Bavisi, Founder and Group CEO, EC-Council. “ADG 2.0 asks for something harder. Say what you do, do what you say, and be able to prove both on any given day. That’s what a regulator, an insurer, or a board will ask for after the next incident. There will be a next incident.”

Controls that run when the agent does

Part of the problem is that agents rarely behave the same way twice. Context shifts, tool responses change, and pre-deployment testing can’t map every path an agent might take once it’s live.

ADG 2.0 places controls at four points in how an AI system operates. The framework calls them ADMIT, DECIDE, EMIT and COMMIT, with human approval required at defined decision thresholds. A tiered autonomy model (Assistive, Conditional or Autonomous) spells out what a system may do at each level and when a person has to step in.

Bavisi draws the line on responsibility by asking who controls what. Model providers secure and monitor the services they run. Enterprises own the deployment: system instructions, connected tools, credentials, data access and the business decisions an agent is allowed to make.

“A provider may see API requests and model-level signals, but it generally cannot see the full business context behind an action,” Bavisi said. “An enterprise knows whether a $50,000 refund is legitimate, whether a payment exceeds delegated authority, or whether a request violates an internal policy.”

Proof that holds up after an incident

Evidence is the other big theme.

Bavisi argues that a model’s stated reasoning shouldn’t be treated as an audit record on its own. Rerunning the same prompt may not help much either, since the model version or the documents it pulled in may have changed. ADG 2.0 defines the evidence each control has to produce, so it gets captured as systems run. Nobody has to piece it together the week before an audit.

For agents, he says, that means a time-stamped record kept inside the company’s own environment. It should show the identity the agent used, its permissions, each material tool call, the policy checks applied and any human approvals or overrides.

“The explanation may help an investigator understand the behavior,” Bavisi said. “The execution record establishes what the system received, what it did, which controls applied and what happened as a result.”

One control set for diverging rulebooks

Fragmented regulation is a compliance headache, and the crosswalks are built for it. Beyond AI-specific rules, they cover security references such as MITRE ATT&CK, the OWASP Top 10 for LLM Applications and the newer OWASP Agentic Top 10.

A company can work from a single control set and show how it satisfies each rulebook it answers to. Regulators get something different: a view of how a proposed national rule lines up with what other jurisdictions already require.

EC-Council is offering the framework and crosswalks free to any government, regulator or standards body. It says it will also help officials who ask to apply ADG 2.0 to their own laws.

“Governments shouldn’t have to build governance from a blank page while they wait,” Bavisi said. “ADG 2.0 is open, it’s already mapped to the rules they’re working with, and we’ll sit down with any government, regulator or standards body that wants help putting it to work.”

ADG is stewarded by EC-Council’s AI Advisory Board, whose members include practitioners from organizations such as Citi, JPMorgan Chase, Microsoft, Salesforce and ServiceNow. The framework also comes with DART (Discover, Analyze, Report, Transform), a method for finding governance gaps and closing them.

Bavisi’s advice for security leaders moving agents into production is practical. Before go-live, put a name next to four jobs: approving the agent’s authority, monitoring it at runtime, suspending it if needed and owning the business outcome.

“You can delegate a task to a probabilistic system,” Bavisi said. “You cannot assume that accountability has been delegated with it.”

The framework, its crosswalks and an AI Readiness Self-Assessment are available free, with no registration, at eccouncil.org/adgframework.

______

About Jay Bavisi

Jay Bavisi is the Founder and Group CEO of EC-Council, the global cybersecurity and AI organization he founded in 2001. EC-Council created the Certified Ethical Hacker (CEH) program and has certified more than 400,000 professionals across 174 countries. Its credentials are recognized by the U.S. Department of Defense under DoD 8140.

 

 

Join our LinkedIn group Information Security Community!

No posts to display