DNS for AI Agent Identity and Discovery

PROPOSED NEW DNS RESOURCE RECORD TYPES FOR AI AGENT DISCOVERY — Researchers propose using DNS, DNSSEC and DANE to help AI agents discover one another, verify identities and establish trusted connections across platforms, with two new DNS resource record types designed for agent discovery. Read more / by Sameer Thakar https://lnkd.in/gp5GX2-R

Our recent IETF Internet-Draft suggests utilizing the existing infrastructure of the #DNS to help #AI agents communicate by introducing two new DNS resource record types. As part of this work, we’ve announced a public license that includes royalty-free terms to support standardization of our proposal. Learn more: https://vrsn.cc/6040BGVPdY

Google Cloud published guidance this week framing agent security as the gating issue for scaling autonomous workflows, and the core recommendation is blunt: do not bolt agents onto legacy access and logging. That sentence describes what almost everyone is currently doing. Agents get service accounts that were created for scripts, inherit permissions nobody has reviewed in years, and write to logs designed for humans clicking buttons. The recommended alternative is specific: task-level provenance, short-lived permissions, end-to-end audit trails, and automatic escalation points. Notice that none of that is about AI. It is identity and access management, a discipline that has existed for decades and that most organizations have quietly deferred maintaining. Agents did not create this debt. They just made it expensive, because a human with stale permissions moves slowly and an agent with stale permissions moves at machine speed. When did your organization last review what permissions its service accounts actually hold? #AI #TechnicalProgramManagement #ProductOperations #DeliveryManagement https://lnkd.in/gUp8wMYr

OpenAI's GPT-6 Astra began rolling out last week, and accompanying it was a security warning from the company itself. Take a look at this week's security developments in AI on our new blog post: https://lnkd.in/giP-V9Y7

BoxLang AI 3.4.0 brings solid improvements. This release introduces gateway support for multiple AI providers, making it easier to switch between services or run them in parallel. The human-in-the-loop feature adds oversight for critical operations, letting you review and approve AI actions before they execute. Security also gets a boost, with a complete stack covering input validation, output sanitization, and rate limiting out of the box. Read the full details here: https://lnkd.in/e8AvG2-Z #BoxLang #AI #Security #DevTools

If we’re willing to allow AI agents to perform human responsibilities, we must also bear responsibility for those agents’ actions as if we’d tasked humans to perform them instead. If you can’t accept the stakes, you can’t afford to sit at the table. Hard to recall a company being quite this reckless at scale in tech but I’m sure finance people could offer some examples. https://lnkd.in/gjpzDfHb

A distinction worth making clearly: there's a real difference between an AI model being trained on your data and an AI model being able to look up your data to answer a question. RunSignup's AI runs on Amazon Bedrock, inside AWS's own infrastructure. Anthropic, OpenAI, and the other model providers never receive your prompts or your event's data — AWS states it directly: "Model invocation communications stay in the AWS network. Model providers can't access the model deployment accounts." Every event's knowledge base is also isolated — queries are scoped to a single event, enforced in code, not just policy. And anything that would create, change, or delete data still needs a person to approve it first. Full breakdown, including what this means for security teams doing diligence on your event: https://lnkd.in/d9Vf_R37 #runsignup #raceday #racemarketing #racepromotion #racemanagement #eventmarketing #eventday #racedirector #registration

Hey folks, Well it was bound to happen. What do you do when ‘tech be techin’. Yesterday, the entire AI industry got humbled at once. On September 3, 2026, four major LLM providers hit trouble within the same window — a good reminder that "the cloud" is still just someone else's servers, and someone else's servers can still have a bad morning. Here's the rundown: OpenAI / ChatGPT & Codex** — The hardest hit. Outage reports spiked past 36,000 on Downdetector, hitting conversations, login, search, file uploads, voice mode, image generation, Deep Research, and Codex. OpenAI confirmed "elevated errors" and rolled out fixes across 15 ChatGPT components and 4 Codex components. Fully resolved by 4:55 PM ET — several hours of disruption. Anthropic / Claude** — Started around 9:30–9:40 AM ET, tied to an infrastructure issue affecting several models (Mythos, Fable, and Opus variants) across claude.ai, the API, Claude Code, and Claude Cowork. Identified and resolved in about 2 hours, with full recovery confirmed around 12:15 PM ET. xAI / Grok** — Confirmed service issues around the same morning window, peaking near 1,365 user reports around 10 AM ET. Shorter-lived than OpenAI's outage, but it hit at the exact same time — which is the interesting part. Google / Gemini** — The least affected, with only around 500 reports by 11 AM ET. Why did they all wobble together? No provider has published a root cause linking all four. It might be coincidence (everyone scaling for a busy week), or a shared upstream dependency (cloud infra, DNS, etc.) — but as of now, that connection is speculation, not confirmed fact. Worth watching for a follow-up statement before anyone draws conclusions. What this means for anyone building on or working with AI tools:** Don't build single points of failure.** If a workflow depends entirely on one LLM provider, you don't have a workflow — you have a liability. Keep a fallback model or provider on standby.** You don't need to run everything twice, but knowing which alternative you'd switch to (and having access already set up) saves you when — not if — this happens again. Save your work as you go.** Download or export in-progress deliverables regularly instead of trusting that a chat thread will always be there when you come back to it. Build in retry logic and error handling** if you're calling these tools via API — treat elevated error rates as an expected event, not an edge case. Check provider status pages directly**, not just Downdetector — official pages usually have more accurate timelines and root-cause updates. One caution on "sharing chats across LLMs but think twice before pasting sensitive information keep a local copy of key context yourself rather than mirroring it into another vendor's system. Bottom line: AI tools are now business infrastructure, not toys. Treat outages like you'd treat a cloud provider going down — because that's exactly what they are. #AI #LLM #Outage #BusinessContinuity #TechNews

Our CISO Jeremy Powell talked AI strategy with TechFinitive. His take: the industry overestimates autonomy and underestimates accountability — every demo shows an agent closing a ticket end to end, but the hard five percent still lands on a human. https://ow.ly/Xz9650ZMI2U

Our CISO Jeremy Powell talked AI strategy with TechFinitive. His take: the industry overestimates autonomy and underestimates accountability — every demo shows an agent closing a ticket end to end, but the hard five percent still lands on a human. https://ow.ly/kz8c50ZNQ2a