Your AI Gateway Is Holding All the Keys
The LiteLLM vulnerability chain is not just another proxy bug. AI gateways sit between users, models, provider keys, stored credentials, and prompt logs. Treat them like production security infrastructure.
Andri Pétur Hafþórsson — security engineer, sysadmin, resident tech wizard in Reykjavík. I keep systems standing under load, under audit, and under a Saturday night.

Eight services, one person, whole stack. Hover a day for the incident log. Unglamorous. Load-bearing. Works.
Fast Next.js, Astro, Shopify and static sites: clear structure, real content, and hosting that stays boring.
Windows, Linux, macOS, DNS, access, backups, logging: infrastructure a small team can operate at 02:00 without crying.
Scope, owners, delivery plans, vendor chasing. Enough process to ship, not enough to become the work.
Information architecture and service flows that make the next action obvious on a phone in bad light.
Architecture reviews, vendor selection, migrations, documentation, untangling systems that grew sideways.
Gap analysis, evidence collection, control mapping, audit readiness, policies written in human.
AI risk reviews, governance notes, logging expectations, operating practices that survive a question.
Threat modeling, disclosure handling, incident readiness, hardening, and very calm triage while it's already on fire.
Fifteen years across systems, security, and audit work. Every major version shipped to production. No rollbacks, several migrations.
Systems administration at the telco that calls itself the biggest party venue in the world. Infrastructure, identity, and the unglamorous load-bearing parts.
Helped customers see their own attack surface the way an attacker does: external attack surface management onboarding, posture, and remediation that actually lands.
Led the VFX studio through TPN Gold, rebuilt the server fleet, planned the new Cisco network, and owned data security through an office move.
Mission-critical air traffic management systems: documentation, on-call, and support for the people keeping aircraft apart. ATSEP-certified.
The whole Microsoft estate for nine branches: networks, a hundred-plus cameras, 3CX telephony, MDM, Power BI, and the procurement to match.
The formative years: where you learn how computers actually break, and how people feel about it.
Built under Kastro, the independent studio. Some live, some parked, one folded — every one of them taught me something.
Client work, useful labs, and independent products from Reykjavík. EU/EEA data residency by default. New client work is paused while I focus on a full-time role; the archive stays public.
End-to-end encrypted messaging with a server-independent core. Works over the internet, LoRa, or a USB stick in a coat pocket.
SPF, DKIM, DMARC, BIMI and report parsing: evidence that your domain isn't trivially spoofable.
Managed home-network safety for Icelandic families. Parked indefinitely: good idea, wrong year.
Iceland-hosted immutable backup, built end to end, then folded into RunAtlas. Shipping it taught me more than keeping it would have.
Security, privacy, and AI tooling — what broke, why, and what actually stops it. Also by newsletter and RSS.
The LiteLLM vulnerability chain is not just another proxy bug. AI gateways sit between users, models, provider keys, stored credentials, and prompt logs. Treat them like production security infrastructure.
KPMG pulled an AI usage report after apparent hallucinations. EFF keeps finding fake staff quoted by AI slop sites. The fix is boring: citations, owners, and a human who checks the receipts.
EDRi says the EU AI Omnibus would delay high-risk AI safeguards and reduce public transparency. That is not boring Brussels plumbing. It is where accountability disappears.
Langflow exploitation is a reminder that low-code AI builders, agent frameworks, and model gateways are not experiments once they touch the internet. They are production attack surface.
Depthfirst says an AI security agent found 21 FFmpeg zero-days for about $1,000. Chrome just patched 429 bugs. The hard part is no longer only finding flaws. It is deciding what gets fixed first.
Active exploitation and AI-assisted attack tooling have compressed patch timelines. Risk-based prioritisation now matters more than a comfortable monthly schedule.
Air navigation, telecom, VFX, universities. Teams where “it's down” is not an acceptable sentence.
No account manager, no discovery phase that bills like a mortgage. The person who does the work is the person who picks up.