BriefTechNews
Poppy 1.0 ships, Infoblox leaks, and Gartner says 40% of AI ops will break by 2028

Poppy 1.0 ships, Infoblox leaks, and Gartner says 40% of AI ops will break by 2028

7 min read · 13 sources

TL;DR
  • Sierra released Poppy 1.0, a protocol for personal agents to interact with businesses, and Shopify enabled it for all merchants.
  • Infoblox is investigating a breach where a compromised migration vendor's access exposed customer network configs and employee data.
  • Gartner predicts 40% of I&O organizations scaling AI will face business-critical disruptions by 2028 due to weak governance.
  • Only 15% of organizations have automatic circuit breakers for runaway agents, despite 95% reporting an agent-related incident in the past year.
  • Amazon lost 40 Kuiper broadband satellites when a ULA Vulcan rocket failed on its first dedicated mission.

Your personal AI agent can now walk into a business the way you would: authenticated, authorized, and with a defined way to get things done. Sierra released version 1.0 of the Personal Agent Protocol, known as Poppy, and Shopify has already flipped the switch for every merchant on its platform. That is the difference between agents calling APIs and agents doing business.

The rest of the day is mostly about what happens when you let agents loose without guardrails. Gartner’s numbers on runaway agents are grim. Cloudflare open-sourced a browser that runs without a browser. And Amazon just lost 40 satellites to a rocket failure. Start with the protocol, because it changes what your agent can do tomorrow.

Only 15% of organizations have automatic circuit breakers to stop runaway agents, despite 95% reporting an agent-related incident in the past year.

Poppy 1.0 gives agents a standard way to transact

Source: sierra.ai ↗

Poppy 1.0 is a protocol that lets personal AI agents interact with businesses through agents, websites, and APIs, with OAuth authentication on top. It builds on MCP, OpenAPI, and UCP, which means it is less a new wire format and more a contract for how agents and businesses introduce themselves, negotiate, and complete transactions.

The v1.0 release adds staged adoption so businesses can roll out support incrementally, plus refined standards usage and a commerce extension co-authored by Shopify. That last part matters: Shopify enabled Poppy for all merchants, and design partners including Delta, Salesforce, and Zoom have joined, with over 100 total. When the largest e-commerce platform ships protocol support by default, the network effect kicks in fast. For engineers, this is the first credible answer to the question of how agents authenticate to your systems without handing over API keys to every LLM wrapper that asks.

Gartner: 40% of I&O orgs will hit AI-driven outages by 2028

Source: gartner.com ↗

Gartner’s warning is blunt: weak governance will cause business-critical disruptions at 40% of infrastructure and operations organizations scaling AI by 2028. The survey behind it is the striking part. 95% of respondents reported an agent-related incident in the past year, yet only 15% have automatic circuit breakers to stop runaway agents.

That gap is the story. Agents fail often, and almost nobody has the mechanical kill switch. A circuit breaker here is not a nice-to-have; it is the difference between an agent burning tokens and an agent deleting a production database. If you are running agents in any capacity, the 15% number is your benchmark. The rest of the industry is one bad prompt away from a pager storm.

Infoblox breach: migration vendor access leaked customer data

Source: theregister.com ↗

Infoblox is investigating a breach where a compromised migration vendor’s access was used to steal SharePoint data. The haul includes customer network configurations and sensitive employee information. Infoblox says it has found no indication that its products, cloud services, or software releases were involved.

Third-party access is the recurring theme in breaches this year. The vendor had legitimate credentials, and someone used them to walk out with data. For anyone running a migration or integration project, this is the reminder that your vendors’ security posture is your security posture. Audit the access you grant, and assume it will be abused.

Cybersecurity at machine speed: the OODA loop problem

Source: x.com ↗

A long read worth your time argues AI is compressing the OODA loop for attackers while defenders stay stuck in institutional governance. Attackers need one successful exploit; defenders must maintain safety, availability, and continuity across everything. That asymmetry is not new, but AI makes the tempo difference structural.

The piece’s point is that adapting means reorganizing decision-making, authority, and governance, not just buying tools. Zero-trust and automated, policy-enforced boundaries are the defensive answer. If you have spent the last year adding AI tooling without changing who can approve what, you have been optimizing the wrong layer.

Cloud overspending is an accountability problem

Source: networkworld.com ↗

Infotech’s Nabeel Sherif argues that cloud waste is not a tooling gap. Flexera’s 2026 report estimates 29% of IaaS/PaaS spend is wasted and organizations run 17% over budget. The cause: operating models that have not caught up with cloud adoption.

Traditional ownership assumptions break in the cloud because teams responsible for costs often lack control, and teams with control lack cost visibility. Sherif recommends redesigning cloud operations around work assignments and team structures. Another dashboard will not fix a team that owns a bill it cannot influence.

Kitesurf is now open source: a browser without a browser

Source: blog.cloudflare.com ↗

Cloudflare open-sourced Kitesurf under Apache 2.0. It is a stateless browser for AI agents that runs entirely on Workers, parsing HTML, applying CSS, executing JavaScript, and rendering screenshots without a browser container. It supports the Chrome DevTools Protocol, so clients like Puppeteer can drive it.

The implementation is Rust compiled to WebAssembly, using packages like Blitz, Stylo, Boa, Linebender, and Krilla. You can deploy it to any Cloudflare account or run it locally. For agent workloads, this removes the heaviest dependency: spinning up and maintaining headless Chrome. If your agents scrape or render pages, this is worth a weekend experiment.

Dell makes enterprise data agent-ready at SC26

Source: dell.com ↗

Dell expanded its AI Data Platform at SC26 with an Agentic Data Gateway, hybrid pipelines, and storage intelligence. The goal is connecting agents to approved enterprise data on customer-controlled infrastructure. Dell also announced direct liquid-cooled storage and a high-density 1U platform expected in the first half of 2027, plus NVIDIA-Certified Storage validation for software-defined PowerScale.

New partner integrations include LangChain, OpenAI Codex, and Unstructured.io. The pattern is clear: every storage vendor wants to be the data layer your agents read from. Dell is betting that agents will need governed access to enterprise data, and that the gateway is the choke point to own.

GitHub Copilot adds Enterprise Server support

Source: github.blog ↗

GitHub Copilot’s latest weekly release adds GitHub Enterprise Server repository support to the Copilot app, now generally available, though cloud sessions are not supported for GHES. Two new models, Claude Fable 5.2 and Gemini 4 Carbon, are available to Pro+, Max, Business, and Enterprise users.

VS Code updates include experimental agent-run model comparisons and an Inbox mode for tracking agent sessions. For teams stuck on GHES, the repo support removes a real friction point, even if cloud sessions remain off the table.

Unit tests still matter when agents write the code

Source: docspring.com ↗

DocSpring analyzed a month of CI history and Claude Code transcripts to test whether unit tests matter when AI agents write code. Useful tests checked decided rules; costly ones were mostly snapshot tests that repeated implementation. One snapshot test was updated 24 times without catching anything.

The key insight: tests describing what code does are meaningless if the agent misunderstood the requirement, because code and test fail in the same way. Tests that enforce explicit rules and run real engines catch critical bugs. The takeaway is not to delete your test suite, but to delete the tests that just mirror the implementation.

Agents need an undo button, not more approval prompts

Source: quaspar.com ↗

This piece argues that approval prompts become reflexive and ineffective over time. The alternative is treating agents as sagas with compensating steps: every tool must have a documented undo mechanism before use. Actions are categorized as reversible, compensable, or irreversible, with the latter needing stricter controls.

The sagas pattern from distributed systems maps cleanly onto agent actions. Tracking and executing undos should live in the underlying system, not inside the agent, because the agent is the thing you do not trust. This is the right framing for anyone building agent infrastructure.

Amazon loses 40 broadband satellites in failed launch

Source: theregister.com ↗

A ULA Vulcan rocket failed to deliver 40 Amazon broadband satellites to orbit after an anomaly during Sunday’s flight. The mission was the first to use a Vulcan configuration tailored for Amazon’s constellation. The impact on subsequent launches remains unclear.

Losing 40 satellites in one shot is a setback for a constellation that needs thousands. The bigger question is whether the failure grounds the Vulcan variant, which would delay the entire schedule. Amazon’s broadband ambitions just got more expensive.

Microsoft's baseline for prompt injection defense

Source: learn.microsoft.com ↗

Microsoft published a defense guide establishing layered prompt injection controls as an enterprise baseline for generative AI systems. The pattern includes prompt shields, spotlighting, plan drift detection, critic agents, tool chain analysis, security guardrails, information flow control, least privilege, short-lived privileges, and human-in-the-loop verification.

The guidance assumes indirect prompt injection will happen and designs accordingly. That is the correct posture: treat prompt injection as a when, not an if, and build the controls to match.

NVIDIA's SkillEvaluator checks your agent skills

Source: github.com ↗

NVIDIA released SkillEvaluator, a framework for evaluating AI agent skills and plugins, installable via pip install skillevaluator for Python 3.12-3.14. It offers quality checks without an LLM key, harness compatibility checks, and CI integration for automated gates before merge. Version 1.0.1 includes a harness-check verifying whether agents like Claude Code, Codex, and Copilot CLI will load a skill.

If you are shipping skills or plugins for agent frameworks, this gives you a CI gate before merge. That is the kind of tooling the agent ecosystem needs more of: not another model, but a way to test the glue.

Get the brief

Liked this one? Pick the beats you care about and get just those, five minutes a morning. Free, no hype, one click to leave.

About Me Author

My name is

BriefTechNews

A daily digest of what actually moved in AI, tech, crypto and fintech, assembled and written with AI, with every item linked to its source. Read More
Tags

You May Also Like