Clint Quah

Payments - Credit Card Network, Web3, Cryptocurrency/Stablecoin Settlements

[Clint Quah @DaddyQuah]

@Jassica445 Hi

[Clint Quah @DaddyQuah]

The Mandiant incident has brought the focus on the emerging Denial-of-Wallet risk, how runaway agents can create security blind spots, and why runtime controls may be essential for keeping autonomous AI within its boundaries.

Can it be stopped when it starts behaving badly?

[Clint Quah @DaddyQuah]

As AI agents gain access to APIs, databases, cloud infrastructure, business systems, and autonomous tools, a software failure is no longer necessarily just a software failure.
Agents can turn an ordinary error into a chain of automated actions with real financial consequences.

[Clint Quah @DaddyQuah]

The AI didn't wait to be attacked. It started probing first.

That is the security signal we should be paying attention to.

New reporting indicates autonomous AI agents were probing Hugging Face networks as early as May—months before the major July incident.

[Clint Quah @DaddyQuah]

Traditional security is built around detecting known indicators of compromise.

Agentic AI changes the equation.

The important question isn't simply what happened in July during the AI agentic bleach.

It's this:
What could have been detected in AI work flow in May?

[Clint Quah @DaddyQuah]

AI security cannot stop at model testing or vulnerability scanning, but continues visibility into what the AI is actually doing at runtime.

The attack may not begin with exploitation.

It may begin with behavior that nobody recognized as dangerous.

#AIRuntimeSecurity ##HexTyx

[Clint Quah @DaddyQuah]

The PaperCut Numbers :

440+ PaperCut deployments compromised.

395 organizations.

48 countries.

5 minutes to domain admin in the fastest confirmed case.

AI isn't just assisting attacks.

It's accelerating them.

[Clint Quah @DaddyQuah]

AI security isn't just about the model.

The real attack surface includes the workflows, tools, APIs, credentials and runtime around it.

RAGFlow shows why.

CVE-2026-45312 and CVE-2026-28797 turned Jinja2 SSTI into remote code execution.

[Clint Quah @DaddyQuah]

Vulnerabilities of AI Part 1 of 7

What if a stranger could borrow your AI workflow���s privileges — without stealing a single credential?

No jailbreak.
No prompt injection.
No malware.
Just a normal question.

That’s Workflow Identity Hijacking.

[Clint Quah @DaddyQuah]

@GioBruno1600 Just one word for Pete, Barry

[Clint Quah @DaddyQuah]

@TrumpsHurricane Had one for 8 years before the great one took the helm. So he ain't the first.

[Clint Quah @DaddyQuah]

AI security isn't just about the model itself. The real attack surface includes

1. The workflows,
2. Tools,
3. APIs,
4. Credentials,
5. Runtime around it.

[Clint Quah @DaddyQuah]

AI-Speed Exploitation

AI is collapsing the gap between vulnerability discovery and exploitation.

The real threat isn't just finding a vulnerability.

It's how quickly AI can turn it into a working attack — and scale it.

[Clint Quah @DaddyQuah]

The HexTyx connection -

This is the thinking behind AIZA-HexTyx — the AI Runtime Immune System:

Don't just ask whether the AI can be attacked.

Test what the AI actually does when it is attacked.

[Clint Quah @DaddyQuah]

The bigger issue -

AI systems don't operate in isolation anymore.

They increasingly have access to:

APIs.
Code.
Credentials.
Cloud infrastructure.
Financial systems.
Customer data.

Autonomy turns a vulnerability into potential action.

[Clint Quah @DaddyQuah]

The uncomfortable question for AI Runtime Security -

The dangerous assumption is:

"Our security controls will never fail."

The better assumption is:

"Our controls will eventually be tested. What happens when they are?"

[Clint Quah @DaddyQuah]

The next generation of AI security won't just protect the model.

It will protect the enterprise from what the model can do.

That's the frontier of AI Runtime Security - https://www hextyx com.

#AISecurity #AgenticAI #AIRuntimeSecurity #Cybersecurity #HexTyx

[Clint Quah @DaddyQuah]

HexTyx — the AI Runtime Immune System will do:

Test the boundary.
Detect unexpected behavior.
Enforce policy.
Block dangerous actions.

Before your AI finds the loopholes of which the DseWiki incident is a warning.

Your AI's real attack surface is what it can actually do.

[Clint Quah @DaddyQuah]

To understand the difference between declared capability and effective capability:

AI agents don't need to be malicious to become dangerous.

They need:
Goal + Autonomy + Tools + Unexpected Capability

That's enough.

Don't trust the boundary. Test it.

[Clint Quah @DaddyQuah]

“Read-only” is not a security control if the agent can turn reading into writing.
The same problem can hit:
→ APIs
→ MCP tools
→ Databases
→ Cloud infrastructure
→ Financial systems
→ Production environments
The real question is not only access but:
What can the agent DO?