DEV Community: Himanshu Vaghela

Shadow AI Is Not Your Real Problem. Shady AI Is.

Himanshu Vaghela — Mon, 24 Aug 2026 14:21:08 +0000

In March, a Meta engineer used an approved AI agent to answer a technical question on an internal forum.

The agent posted its answer publicly. Without approval. The engineer followed the advice. Two hours later, a large volume of sensitive company and user data was sitting in front of employees who were never authorized to see it. Sev 1.

Here is the part most people miss: this was not shadow AI. The tool was vetted. It was approved. It was rolled out. Nothing was rogue and nothing was smuggled in.

The tool was fine. The approval was the problem.

A new category is forming. Shadow AI is the unapproved tool you cannot see. Shady AI is the approved tool doing something nobody anticipated, in plain sight, inside your own perimeter. And you cannot block it, because you already said yes to it.

A July 2026 survey found 76% of security teams now carry some responsibility for governing enterprise AI. But the controls are still built for tools, not for actions. We approve which software is allowed in. We rarely verify what that software does in the moment it acts.

The agent did not break in. It was invited, given a seat, and nobody was standing at the door checking what it was about to hand over.

So the question is no longer whether the model was approved.

It is whether anything verified what it did before the data was already out the door.

Documentation Is Not Governance. The EU AI Act Knows the Difference.| Himanshu Vaghela

Himanshu Vaghela — Sat, 01 Aug 2026 08:02:41 +0000

2nd August 2026.

That's the EU AI Act deadline most organisations are sprinting toward.

Here's what nobody is saying out loud:

Most teams will hit the deadline with documentation.

Not governance.

There's a difference.

Documentation says: "We have a human in the loop."

Governance proves: "Here is what the AI was authorised to decide. Here is
what it actually decided. Here is the evidence that existed at the moment
it acted."

Article 14 of the EU AI Act requires "meaningful human oversight".

Meaningful is doing a lot of work in that sentence.

A human who clicks approve in 3 seconds without context is not meaningful oversight.

A human who cannot stop the workflow before consequence is not in the loop.

A policy document that describes what should happen is not evidence of what did happen.

73% of production AI deployments are vulnerable to prompt injection.

Only 11% of organisations have implemented governance frameworks for AI agents.

Both statistics are from OWASP 2026.

Both will matter to regulators on 3rd August.

The organisations that survive regulatory scrutiny will not be the ones with the best policy documents.

They will be the ones who can open a workflow from last month and show:

What the AI was authorised to decide.
What it actually decided.
Who had authority to approve it.
What evidence existed before it acted.

That is not a compliance exercise.

That is verification architecture.

And most teams haven't built it yet.

2nd August is not the finish line.

It's the starting gun for enforcement.

Himanshu Vaghela
Junior Researcher, LumiRosh Research.

The Human in the Loop Fallacy: Why 'Someone Will Check' Is Not a Control

Himanshu Vaghela — Mon, 27 Jul 2026 19:30:04 +0000

There is a moment in every AI deployment conversation that should make security people flinch.

Someone says: "Don't worry, there's a human in the loop."

And everyone nods.

Let me tell you why that nod is dangerous.

What "Human in the Loop" Actually Means

Let's look at how AI output becomes business action in most organisations today:

LLM generates a summary of a customer support ticket
Human reads it quickly between meetings
Human clicks "send" because it looks fine
Output becomes action, the customer gets that response

Was there a human in the loop? Yes. Was there verification? Barely.

The gap between "a human saw this" and "a human verified this with the right context, authority, and accountability" is where most AI risk lives.

The Three Illusions of "Human in the Loop"

Illusion 1: Presence = Verification

Just because a human sees AI output doesn't mean they verified it. Research shows that humans in AI assisted workflows:

Accept AI suggestions 80%+ of the time (automation bias)
Spend under 3 seconds reviewing AI generated content before acting
Are more likely to approve output when it's well formatted, even if factually wrong

The control is not working if the human isn't really checking.

Illusion 2: Familiarity = Authority

The person reviewing AI output is rarely the person who should be making the decision.

In practice:

A junior support agent reviews an AI drafted escalation recommendation
A developer approves an AI generated code change their manager won't see
A compliance officer signs off on AI drafted regulatory filings

The "human in the loop" is often the wrong human, present but not authorised.

Presence without authority is theatre, not control.

Illusion 3: Retrospective Oversight = Prevention

Some teams interpret "human in the loop" as: "We'll audit AI decisions after they happen."

This is not a control. This is a post mortem.

When AI output has already triggered a payment, sent a regulatory communication, or modified a production configuration, reviewing it afterwards doesn't undo the action.

If you can only detect the failure after it executes, you don't have a loop. You have a log.

What Actually Works: Verification Before Execution

The difference between human in the loop and human verification is simple:

Dimension	Human in the Loop	Human Verification
When	During or after	Before execution
What's checked	Output looks fine	Claims are true, authority is valid, scope is correct
Who checks	Available human	Authorised human with context
Evidence	None kept	Verifiable record
Can fail?	Yes, by design	Failures are blocked

In an AI verification system:

AI produces output, a recommendation, a draft, an analysis
Verification checks execute, is the source cited real? Does the human have authority for this action? Is the scope within policy?
Human reviews with context, not just the output, but the verification findings
Human makes an informed decision, approve, reject, or escalate

This is what we call Execution Assurance at LumiRosh. The output is not the evidence. The verification before execution is the evidence.

Why This Matters Now

We are in a period where AI deployment is accelerating faster than control design.

73% of production AI deployments are vulnerable to prompt injection, and only 11% of organisations have implemented governance frameworks for AI agents (OWASP 2026)
Automated AI decisions are doubling every 6 months
Regulatory pressure (EU AI Act Article 14) now mandates meaningful human oversight

The organisations that treat "human in the loop" as a checkbox will be the ones discovering compliance gaps the hard way.

The ones that build real verification will have the audit trail to prove it.

The Bottom Line

If your AI deployment has a human somewhere in the vicinity of the output, you do not have a control.

You have a false sense of one.

Verification is a design problem, not a seating chart. If the human is not equipped, authorised, and required to check before action, they are not in the loop. They are in the way.

I work on AI Execution Verification at LumiRosh Research, mapping where output becomes action, and where human judgment needs a stronger control point.

Connect:
linkedin.com/in/himanshu-vaghela9

The Architecture Of Execution Assurance

Himanshu Vaghela — Fri, 19 Jun 2026 13:30:50 +0000

A team deploys AI into a real workflow.

Output starts flowing. Speed increases. Everyone feels progress.

Then something breaks. A decision was made but no one can explain why. The AI approved an expense, rejected a candidate, or escalated a ticket. And when someone asks "why did it do that?" the answer is silence.

This is unverified execution. And it is the single biggest risk in AI adoption today.

At LumiRosh, we built the AI Execution X-Ray a 5-day forensic diagnostic that answers three questions:

What is your AI actually deciding?
Can you verify those decisions?
Is there an audit trail?

The answer for most companies: they have speed, but zero accountability.

Here is what we have found across 30+ organizations:

The Pattern
Every team follows the same arc. They start with a pilot one workflow, one AI tool. The output looks good. They expand. Within weeks, AI is embedded in decisions they never intended. No governance. No verification layer. Just trust.

And trust is not an audit trail.

The Fix
Execution Assurance is not about slowing down. It is about building a verification layer into the workflow a zero-trust checkpoint that says: every AI decision must be reproducible, verifiable, and attributable.

No governance theater. No compliance paperwork. Just evidence.

The Result
Companies that implement Execution Assurance catch hallucinations before they reach the customer. They sleep better at night because they know not hope that their AI is doing what it was told.

We wrote the full Architecture of Execution Assurance on Medium. Link in bio.

Find me at
Himanshu Vaghela or LinkedIn.