29. Security Of, From and Using AI

While I worked at Google during the publishing of this post / video, the views expressed here are my own and may not reflect those of my employer. Only publicly available material is used to put together the content of this article. The website and the Youtube videos are NOT monetized.

Many media outlets called this month the Black October due to various cybersecurity issues and attacks discovered. Regardless of what you call it¹, I am glad there seems to be a consensus that security from and more importantly using AI needs to be at the forefront.

Before I lose your attention, let me provide you everything you need to know (and act) in the next few (less than 100) lines.

Intersections of AI + Security

The intersection of Security and AI is at three points. Security Of AI, From AI and Using AI.

Security Of AI - Securing your own AI (training data, pipelines, model weights, endpoints, AI-enabled applications and agents) and your AI supply chain against at least the Top 10 LLM and top 10 Agent attacks.

Security From AI - Securing your (and or your customers') access, infrastructure, data, and applications (AI apps or otherwise) from adversaries that are using AI-enabled attacks at an unprecedented scale.

Security Using AI - Securing against traditional and AI-enabled threat actors using AI by being extremely proactive and deliberate about it.

Let’s be on the same page

Before I give you what you can do next, here are two ground realities from my over a decade in Cybersecurity and Sovereignty (1, 2) and recent few years of AI + Security work:

1. The Dual-Stack Burden: For the foreseeable future, enterprises and governments still have to operate and defend their traditional stack (core banking, citizen portals, email, chat, SaaS, legacy databases). AI (models, MCP servers, autonomous agents, RAG pipelines) is additive—you haven’t replaced the old attack surface; you’ve expanded it! You have to protect both.
2. The 3-Way Threat Multiplier:
   • Speed: Traditional threats now run at AI velocity—Mean Time to Exploit (MTTE) has collapsed from weeks to hours.
   • The Insider Agent: You must protect your AI from turning on you (prompt injections, unauthorized tool execution, silent data exfiltration).
   • Outgoing Liability: If your agent or exposed toolchain is hijacked to attack partner networks, customers, or critical infrastructure, you carry the regulatory, sovereign, and financial liability.

The 20% effort that gives 80% back (and beyond)

Note - Medium and large enterprises should do everything that small enterprises are advised plus what’s advised specially to them

1. Security using AI

I moved this above others because

1. Using AI is the only way to protect yourself against AI-enabled threats at the AI scale
2. AI-augmented defnese/prevention protects both AI and non-AI applications and infrastrcuture against both legacy and modern threats.

Do this in short term -

Do this in the long term - Build on the foundations that you created in the short term

1. Building a map/knowledge graph of entities (infrastructure items, applications, services and identities) and their relationships (reachability, IAM access). Products like Wiz Cloud provide this out of the box with Wiz Security graph. Hugely influential in figuring out the blast radius and priority when needed.
2. Continuous AI Assisted Automated Red-Teaming (CART) for production systems, API contracts, and IAM policies 24/7/365, reporting vulnerabilities before external actors find them. Proactively and Deliberately do what your adversary would do but with more internal context than the attacker - you will always be ahead in finding and fixing holes in your system.
3. Develop an AI-assisted Scan - Remediate - Monitor loop (check AI Threat Defense)

2. Security from/against AI

Quick Note - This includes defending against your own (compromised) AI as well.

Do this in short term -

Do this in the long term -

3. Security of AI

Quick reminder: The risk is rarely to the model or weights in isolation—the real risk is to your data, your customers, your reputation, and your infrastructure being weaponized against others via your exposed AI endpoints, tools, and agents. Also it is assumed that your impeccable traditional security (mentioned in the section above) has done the heavy lifting to secure your / your customers data that might be used for training your AI / by your AI (e.g. RAG applications)

Do this in short term -

Do this in the long term -

Conclusion

The Bad

1. Adversaries have AI now: The barrier to entry for cybercrime has collapsed. Automated reconnaissance, instant vulnerability fuzzing, and hyper-personalized spear-phishing mean anyone can hack, hack faster, and attack anyone at machine speed.
2. Ignorance is no longer an excuse: Courts, regulators, and your customers will not accept “the AI did it.” You and your leadership are on the hook for:
   • Unauthenticated or Unattributed AI Access: Exposing model endpoints, agent runtimes, or MCP servers without strict authentication, or letting autonomous agents interact with production data and external APIs without verifying the real human identity legally behind the request.
   • Zero Observability and Governance: Failing to log, monitor, and define explicit policy boundaries (who is allowed to do what—and who is blocked) across the entire chain: Human $\leftrightarrow$ Agent $\leftrightarrow$ Tools (MCPs, APIs) $\leftrightarrow$ LLM.
   • Passing the Blame for Your AI’s Actions: Blaming an autonomous agent for an outage, a data breach, or an unauthorized transaction does not hold up in court or with regulators. Your AI is your infrastructure; you own its liabilities.
   • Deploying What You Don’t Understand: Shipping, using, or providing anything to customers—whether code, automated workflows, or critical decisions—that nobody on your team understands because “the AI generated it."

The Good

1. Traditional, scalable security solves the vast majority of threats: The fundamentals—Safe Code (with SAST), mature vulnerability management (for infra, PCP and SaaS configurations and code) least privilege, zero trust, rapid patching, network segmentation, and phishing-resistant MFA—are proven, largely solved engineering practices. They were often deprioritized in the past because “not everyone had the skill to hack.” Now that AI democratizes exploitation, rigorous security hygiene is no longer optional; it is your baseline armor.
2. You can fight AI with AI (The Defender’s Advantage): Asymmetry does not have to favor the attacker. You have an insurmountable structural advantage that external adversaries will never have: deep internal visibility, ground-truth context, and proprietary telemetry into your own architecture. When you pair proactive, continuous AI defense with your internal knowledge graph, you can detect, contain, and remediate threats long before an adversary maps your perimeter.

The Everyday

1. Watch out for security bulletins for attacks, threat actors, and vulnerabilities. If you have a system that tells you whether a bulletin applies to you, what the blast radius is, and how to fix it—you are already in great shape, especially if this is automated and AI-assisted!

Thank you for reading through!

Please share 🔗 if you think it’s useful! And please comment ✍ if you want to discuss more!

— Nikhil

Further Reading & Resources

Here are the essential frameworks and deep dives across each dimension:

Acronyms & Terminology Glossary