Phalgun

CEO & Co-founder, Verigrey, I mainly write about agent behavioral assurance and security.
Verigrey found 90% more agent violations in benchmarking.

[Phalgun @phalgun]

Deploying Agents in your enterprise? We are solving agent trust at https://t.co/vVGa0GRJbc - we've built the Behavioral Assurance Platform for AI Agents. Not just security, but the entire scope of potential rogue behaviors such as compliance, reliability, security, and more, so teams can deploy agents they trust before launch and monitor after. We found 9x more violations compared to typical benchmark approaches. #AI #Assurance #cybersecuity #agents

[Phalgun @phalgun]

Most AI agent testing today is manual or black-box: send prompts, read responses, hope you covered enough.

That's not enough when agents take real actions on behalf of banks, insurers and governments.

Today Verigrey joins the AI Verify Foundation to help change that. Our cutting-edge assurance technology, looks inside how an agent actually behaves — finding security, compliance and functional violations before customers or regulators do and continuously monitors and controls behaviors in production.

Proud to be part of the mission to make AI testable and trustworthy.

#AI #AIagent #AgenticAI #CyberSecurity #AIGovernance #Compliance

[Phalgun @phalgun]

Glad to see AI safety is paramount in this expansion to Singapore. Singapore is leading most the world in ensuring independent safety testing of agents.

[Anthropic @AnthropicAI]

We’ve opened an office in Singapore—our fifth in the Asia-Pacific region. We’ve also signed an MOU with the Government of Singapore on AI safety and responsible adoption.

You can read more about our new office here: https://t.co/125mBgHLkM

[Phalgun @phalgun]

@hunvreus Signed up - looks fun!

[Phalgun @phalgun]

They took many months to test after building. Right now enterprises take weeks or months to approve an agent with manual testing that is much less effective in finding agent behavioral issues. This is why we built Verigrey - giving proof for enterprises to deploy agents that behave safely in 1-2 hours - solving a big piece of the missing puzzle.

[Phalgun @phalgun]

@alexandr_wang Right now enterprises take weeks or months to approve an agent with manual testing that is much less effective in finding violations. This is why we built Verigrey - to do agent behavioral assurance in 1-2 hours pre-deployment - solving a big piece of the missing puzzle.

[Phalgun @phalgun]

This is why we built Verigrey - to do agent behavioral assurance in 1-2 hours pre-deployment. This includes testing rogue behaviors, security and policy compliance. Right now enterprises take weeks or months to approve an agent with manual testing that is much less effective in finding violations.

[Phalgun @phalgun]

When your agents discover they can communicate in ways you never tested for.

Test agent behavior with https://t.co/KUUapqQE0Y

[Phalgun @phalgun]

3. Your biggest agent risk may be the agents you don’t know about

Employees are already connecting AI to enterprise data and tools. Vendors are embedding agents inside their products. Business units can deploy agents faster than traditional security reviews were designed to handle.

The result could be the AI equivalent of shadow IT — except this software can act.

CISOs need visibility into what agents exist, what they’re connected to, what permissions they have, and what they’re actually doing.

These aren’t just new vulnerabilities.

Agents change the security model itself.

When software can reason, use tools and take actions at machine speed, we need security and assurance that operate at machine speed too.

[Phalgun @phalgun]

Three worries every CISO/AI Steering Committee should have about AI agents

1. Agents are becoming privileged users

Agents aren’t just generating text anymore. They’re getting access to APIs, databases, browsers, internal systems and other agents — and taking actions autonomously.

We spent decades building identity, access controls and monitoring around humans.

Now we’re creating a new class of machine user that can act thousands of times faster.

CISOs need to ask: Do we actually know what our agents can access, what actions they can take, and whether they stay within those boundaries?

🧵 Below 👇

[Phalgun @phalgun]

2. A safe agent can become unsafe without being hacked

Cybersecurity traditionally focuses on compromise: Was an attacker able to get in?

With agents, that’s no longer the only question.

An uncompromised agent can still take an unintended action, misuse a legitimate tool, expose sensitive information, violate policy or behave differently when its environment changes.

That means security has to expand from protecting the system to continuously assuring its behavior.

🧵👇

[Phalgun @phalgun]

I think #1 leads to an important shift: from AI security to AI assurance.

Traditional security asks: Has the system been compromised?

Agentic systems add another question: Is the agent actually behaving as intended?

When software can reason, use tools and take actions at machine speed, enterprises will need both answers continuously.

[Phalgun @phalgun]

😅😂😂🤣

[Greg Hunkins @greghunkins]

Good Will Hunting GPU Edition
Ep. 2: My Boy's Wicked Smaht

[Greg Hunkins @greghunkins]

Good Will Hunting GPU Edition
Ep. 2: My Boy's Wicked Smaht

[Phalgun @phalgun]

Can find MAS' Safeguards for Agentic Finance at Runtime (SAFR) report here: https://t.co/yAp2sSbCBf

[Phalgun @phalgun]

Singapore may be quietly building one of the world’s most complete governance frameworks for AI agents.

In July, MAS and HSBC, JPMorgan, Mastercard, Visa, OCBC and others published SAFR—a framework for governing agent actions at runtime.

Now MAS’s AI Risk Management Guidelines explicitly call for testing agent failure modes and guardrails, independent validation of high-risk AI, and ongoing monitoring.

The emerging model is becoming clear:

Test before deployment. Govern every action. Monitor continuously. Maintain evidence.
That’s what agent assurance is going to look like.
This is exactly what we purpose built Verigrey to solve - rigorous behavioral assurance of AI agents.

[Phalgun @phalgun]

@a16z Correct outcomes ≠ correct behavior. This is what most enterprises are grappling with their agents today. It is not just a point security issue, it is behavioral. That what we built https://t.co/KUUapqQE0Y to solve - behavioral assurance.

[Phalgun @phalgun]

This is the exact problem we are solving at https://t.co/vVGa0GSh0K - we've built the Behavioral Assurance Platform for AI Agents. Not just security, but the entire scope of potential rogue behaviors such as compliance, reliability, security, and more so teams can deploy agents they trust before launch. @VerigreyHQ

[Phalgun @phalgun]

This is the exact problem we are solving at https://t.co/vVGa0GSh0K - we've built the Behavioral Assurance Platform for AI Agents. Not just security, but the entire scope of potential rogue behaviors such as compliance, reliability, security, and more so teams can deploy agents they trust before launch. #AI #Assurance #cybersecuity