What changed in AI security, and what it means in the interview.

The three identity problems agents create (acting for a user, proving themselves to websites, and being governed as a fleet), and the mechanisms that answer each in 2026: delegated OAuth tokens, Web Bot Auth signatures, agent registries and gateways.

How AI agents pay without holding card numbers: signed mandates for user consent, scoped payment tokens, signed agent requests for merchants, and the liability questions that remain open in 2026.

How AI vulnerability-discovery and auto-remediation systems work in 2026, what Google, Anthropic, OpenAI and Microsoft have shipped, why validation is the hard part, and the interview questions this now produces.

How browser extensions get access to every session, how developer phishing and ownership changes turn trusted extensions malicious, what the Cyberhaven incident showed, and the Chrome enterprise controls that limit the damage.

How attackers abuse the OAuth device authorization flow to get tokens after a victim signs in on the genuine Microsoft or Google page, why passkeys do not stop it, and the Conditional Access policy that does.

What fail2ban and CrowdSec do, why log-driven IP banning is not authentication, and the SSH design (keys or certificates, no passwords, no internet exposure) that makes brute force irrelevant.

What sender authentication proves and what it cannot, how secure email gateways and API-based filters differ, why links are checked at click time, how QR-code phishing gets around controls, and what happens after a phish lands.

A clear explanation of mesh VPNs such as Tailscale, headscale and NetBird: key distribution through a coordination server, direct WireGuard tunnels, NAT traversal and relays, and the security questions interviewers ask about them.

What infostealer malware takes, how stealer logs are sold, why stolen session cookies bypass MFA, what Snowflake 2024 and the Lumma takedown showed, and the controls that make stolen secrets worthless.

A sober look at malware that calls large language models at run time: LameHug and PROMPTSTEAL, PROMPTFLUX, the PromptLock proof of concept, supply-chain malware abusing local AI tools, and the detections that work.

How operational technology differs from IT, what IEC 62443 zones and conduits mean, which real incidents shaped current guidance, and the controls that work when equipment cannot be patched for months.

The patch and vulnerability-response questions security interviews ask, with the mechanics behind strong answers: deployment rings, installed versus running code, containers, edge appliances, patch diffing and incident-style CVE response.

Why apps generated by AI app builders so often expose every user's data, what Row Level Security does, which keys are safe in a browser, what CVE-2025-48757 showed, and a one-hour review any team can run.

What Wiz-style cloud scanning, SASE, bot management, AI security platforms, data security posture management and EDR actually inspect and enforce, with the trade-offs interviewers expect you to know.

A security-focused guide to US AI law as of November 2026: what California SB 53 requires, how New York's amended RAISE Act differs, why Colorado's AI Act never took effect, what Texas TRAIGA prohibits, and what the federal executive order does and does not do.

Five documented AI security incidents, the boundaries that failed, and how to choose controls for infrastructure, model files, caches and agent actions.

What interviewers want when they ask about certificate pinning: what it stops, what to pin, how to rotate without outages, how testers bypass it on their own devices, and why the API must be secure anyway.

Internal compromises rarely need new exploits. They follow credentials from machine to machine. This post walks the usual Active Directory attack path step by step and pairs each step with the control that breaks it.

An adaptable AI security study plan with practical exit criteria, a worked agent-design exercise and a method for reviewing your own answers.

What a VPN actually protects, how IPsec's IKE and ESP compare with WireGuard's cryptokey routing, why NAT matters, and why the gateway and the access policy behind it decide most VPN incidents.

Most Linux privilege escalation comes from configuration, not clever exploits: writable files run by root, permissive sudo rules, setuid programs and stale kernels. Here is how each works and the controls that close it.

Most serious mobile findings do not need a rooted device. Static analysis, IPC and deep link testing, direct API testing, test builds and repackaging cover most of an assessment, and the gaps should be reported honestly.

How mutual TLS works, what it proves and what it does not, where termination at proxies changes the trust, and the follow-up questions interviewers use to separate memorised answers from real understanding.

Apply the lethal trifecta to a coding assistant, find hidden outbound routes, and turn an initial triage into a specific control and a test.

A seven-stage order for learning the fundamentals security interviews test, from how networks and operating systems work to TLS, platforms, application flaws, attacks, AI systems and incident response, with what to practise at each stage.