Security leaders are making AI security decisions with almost no data about what their peers are doing. Sounil Yu and I built a five-minute anonymous survey to change that, and we'll publish the findings so you can benchmark your approach. https://zeltser.com/ai-security-survey

I've seen too many author-centric assessment reports, full of stories of conquest and irrelevant details. The best reports are reader-centered. My new template offers a consistent way to write them that way. It's AI-friendly, too. https://zeltser.com/security-assessment-report-template

A decade of CA failures could have collapsed the web's certificate trust model, but they didn't. Browsers and CAs addressed each failure with a structural fix, and the same approach applies wherever you delegate trust. https://zeltser.com/past-present-future-web-trust-model

Sounil Yu and I built the AI Defense Matrix, an approach for deciding how to defend the various AI systems in your environment. Here's why. https://www.sans.org/blog/why-sounil-yu-i-built-ai-defense-matrix

AI capabilities of keyboard apps (rewrites, voice transcription, generative writing) entice us into letting our keystrokes float to the developer's servers. That makes each one a keylogger we authorized, with safeguards left to the developer. https://zeltser.com/third-party-keyboards-security

My new template for security assessment reports offers a structured, repeatable approach to communicating with readers. It's based on my experience creating and reading many such reports over the years. It's AI-friendly, too. https://zeltser.com/security-assessment-report-template

How can you use your AI agent to transform raw threat notes into a well-sourced CTI report draft? Use my MCP server with my new CTI report template and writing guidance. Your sensitive data stays local. https://zeltser.com/cyber-threat-intel-report-template

Observability, short-lived credentials, and active enforcement held the web's certificate trust model through a decade of CA failures. The same three levers work anywhere you delegate trust, from code signing to identity federation. https://zeltser.com/past-present-future-web-trust-model

My new template for cyber threat intelligence reports covers tactical, operational, and strategic aspects of threat activity. A companion brief captures the key takeaways for decision-makers. You can also use it with your AI agent. https://zeltser.com/cyber-threat-intel-report-template

Threat attribution works at 3 levels: Tactical examines the incident, operational characterizes the campaign, and strategic asks who's responsible and why. Disciplined analysts weigh the same 6 signals at every level. https://zeltser.com/six-signals-for-threat-attribution

Every domain has its Turbo Encabulator. The more expert we feel, the more fluently we explain something that doesn't exist. https://zeltser.com/experts-overstate-expertise

This template for a security incident report not only helps with documentation, but also offers guidance when capturing findings during the incident. Take it, customize it, use it. https://zeltser.com/incident-response-report-template

Plant a decoy persona, such as a fake LinkedIn profile or an unused directory account, to catch attackers reaching for what they think is real. The decoy has no production use, so any interaction is a high-confidence alert. https://zeltser.com/the-notion-of-a-honeypot-persona

We read more threat attribution claims than we make. Six signals separate the ones that hold up from the ones that don't, and analysts weigh them together to build a defensible case. https://zeltser.com/six-signals-for-threat-attribution

Letter grades tied to specific criteria influence business' behaviors, as NYC's sanitation labels have shown. Cybersecurity is starting to follow with the Cyber Trust Mark for IoT, but the consumer web hasn't caught up. https://zeltser.com/restaurant-inspections-as-security-validation-model

A decoy persona alerts you when an attacker probes your directory or public profiles. When they can't tell the bait from real identities, you create an asymmetry that gives you the upper hand. https://zeltser.com/the-notion-of-a-honeypot-persona

"The career-shaping work happens in the years before the job search." 100% We're seeing others say the same things albeit in other areas, such as leadership

The career-shaping work happens in the years before the job search. Reputation, relationships, and visible work are what land the right role. The resume's job ends at initial screening. https://zeltser.com/stop-relying-on-your-resume

Having interviewed many candidates, I can tell within minutes whether someone prepared for the discussion. A few hours of focused prep often makes the difference between getting the offer and getting a polite no. https://zeltser.com/interviewing-tips-for-it-job-candidates

The REMnux MCP server can now draft malware analysis reports using my new report template: https://zeltser.com/ai-malware-analysis-remnux

Self-hosted Algo on DigitalOcean lets us treat the VPN exit IP as disposable. After investigating malicious infrastructure, destroy the droplet, redeploy in minutes, and the next project starts from a clean IP. https://zeltser.com/deploy-algo-vpn-digital-ocean

I updated my cheat sheet for creating cybersecurity assessment reports. It's a one-page doc, which you can customize: https://zeltser.com/security-assessment-report-cheat-sheet

The people who handle breaches all day may be the worst at protecting themselves. Feeling invulnerable is what lets us function around constant threat, the way it lets doctors work around disease. Warnings about our own risk rarely stick. https://zeltser.com/illusion-of-invulnerability

A decoy fires only when someone accesses a resource no legitimate user would touch. Plant tripwires across network, identity, data, and AI agent configs to create asymmetry in your security architecture. https://zeltser.com/protean-information-security-architecture

One word changed a hospital hand-washing sign from 'protects you' to 'protects patients,' and compliance climbed. We discount our own risk but not other people's. The same holds for security messaging aimed at others, not ourselves. https://zeltser.com/illusion-of-invulnerability

An attacker on a developer's machine often pivots to reconnaissance. AI agent MCP configs are plain-text files at known paths, offering an index of high-value services. A decoy entry pointing to a honeypot MCP server alerts you of an intrusion. https://zeltser.com/decoy-mcp-server-honeypot

One should print out those rules, laminate them and give it to everyone in cyber.

When an executive rejects a security recommendation, it's worth asking what would need to change for a different answer. That question reveals constraints we didn't see and persuasion paths we didn't consider. https://zeltser.com/rejected-security-recommendations

"Fileless malware" started as Code Red's pure in-memory worm in 2001 and has evolved to cover other forms of evasion, including living-off-the-land. I traced the journey of the term through its 25-year history. https://zeltser.com/fileless-malware-beyond-buzzword

Much of our security work is communicating with colleagues throughout the org. 10 habits that sharpen how the technical work gets heard. https://zeltser.com/strong-communication-skills

Four tips for a strong executive summary of your security report. They apply whether you write it from scratch or draft it with AI's help: https://zeltser.com/executive-summary-for-security-assessment-report-tips

Existing AI security frameworks each cover one slice of the work, like components, risks, and lifecycle. The AI Defense Matrix combines them into a single grid of AI asset classes mapped to NIST CSF functions. https://zeltser.com/ai-defense-matrix-intro

A honeytoken fires when someone reaches for what they shouldn't, which makes the alerts high-signal. To get the most of them, decide on the best locations, such as decoy MCP entries, fake AWS keys, and Cloudflare Workers serving fake admin pages. https://zeltser.com/plant-honeytokens

If AI driven attacks become more prevalent, it'll only be a matter of time before attackers push the token burden on to their victims, using the AI that's already (probably) there. I'm calling it "living off the lAInd".* *Jokey name. Probably will happen, though.

A decoy in your AI agent's MCP config can be an early sign of an intrusion. Interactions with the honeypot MCP server mentioned there can be a high-confidence signal. Building this honeypot is pretty straightforward. https://zeltser.com/decoy-mcp-server-honeypot

Sounil Yu and I co-authored the AI Defense Matrix, the security-for-AI companion to his Cyber Defense Matrix. It maps eight AI asset classes to NIST CSF functions, so security leaders can find gaps and vendors map products. https://zeltser.com/ai-defense-matrix-intro

What exactly is "malware"? Here's my definition, so we don't need to rely on "We know it when we see it." https://zeltser.com/what-is-malware

Unnecessary complexity makes products hard to maintain and hard to secure. Modern apps such as Cloudflare's EmDash and Tailscale show that designing for simplicity produces stronger security as a side effect. https://zeltser.com/modern-design-security

Honeytokens are a time-tested idea, but the interesting part is where you plant them. Consider using them as decoy MCP entries, fake AWS keys, and Cloudflare Workers serving fake admin pages to detect intrusions. https://zeltser.com/plant-honeytokens

How do builders of security products assess strategies against vibe-coded competition? You can now use your AI agent and my MCP server, which carries Ben Vierck's seven-dimension defensibility rubric, included with his permission. https://zeltser.com/security-product-strategy-with-ai

SaaS vendors should assess whether their trust boundary includes customers' AI agents. Liability has pushed banks toward securing the customer's device four times, and the fifth wave is forming around AI agents. https://zeltser.com/saas-ai-agent-trust-boundary

Modern architectures make products easier to run, with security as a wonderful added benefit. Cloudflare's EmDash reimagines WordPress with no customer-managed server, and Tailscale connects devices with no VPN servers. Simpler designs leave less to attack. https://zeltser.com/modern-design-security

When evaluating a security vendor's strategy, your own product roadmap, or an investment target, AI can separate marketing claims from verified capabilities. But it needs the right domain-specific criteria. Here's how: https://zeltser.com/security-product-strategy-with-ai

Generic AI does generic work. Once Claude Code knows your tools, your conventions, and your past projects, its outcomes start fitting how you actually operate. The seven-layer Personal AI Stack lays out what to add and why. https://zeltser.com/personal-ai-stack

AIUC-1 is a purpose-built compliance framework for AI agent risks such as prompt injection that existing certifications don't cover. Scope, auditor dynamics, and incentive alignment will shape what its certificates are worth. https://zeltser.com/aiuc-1-cert

Out-of-the-box Claude Code is solid for general work. What makes it indispensable is the personalization, hardening, and connectors you layer on top. I mapped my setup into a seven-layer Personal AI Stack, so you can optimize yours.

SaaS vendors that make their products usable by customers' AI agents inherit those agents' attack surface. Liability and regulation drove banks into that position four times already. The same pressure is building for agent-era vendors now. https://zeltser.com/saas-ai-agent-trust-boundary

"The current generation of frontier models behaves like a gifted PhD student with imposter syndrome: brilliant when calm, and when the room turns against them, they over-apologize, hedge everything, and abandon positions they should defend." -- Dheer Gupta https://dheer.co/llm-anxiety/

Dan Nguyen-Huu walks through the economics of agentic procurement, a scenario where AI agents decide whether to build a capability from scratch or buy it from a vendor. https://dannguyenhuu.substack.com/p/the-token-threshold

You can now receive my blog posts via email. Go ahead and sign up: https://zeltser.com/newsletter I've been writing more frequently than I have in recent years, and I'd rather share my articles directly than rely on an algorithm to decide whether to surface them.