iiinova — where your digital self keeps evolving

Claude went down this morning: chat, Code, Cowork, and the API. Anthropic said do not sign out. 9to5Google counts five incidents this month. If your agents are pinned to one model, your failover plan is a status page. What runs when the model does not?

000

Axonius and AWS announced a strategic collaboration today to ground the AI agent in customer asset data and isolate it inside Bedrock. The premise: the agent is only as reliable as the inventory it can see. IT spent decades building the CMDB. Agents gave everyone a reason to keep it current.

000

AWS spent nine months interviewing 154 executives for Reimagine 2026. The line that carries the week: CIOs who fought shadow IT for years now face shadow AI at ten times the scale. A two-week experiment that waits a month for approval stops asking. Which half of your AI estate did you approve?

011

NVIDIA's OpenShell got a data layer today. Bedrock Data plugged Agent DLP into the sandbox: every action is judged against the data involved, not the permission granted. Permissions can be perfectly scoped and the leak still happens. Least privilege was never data-aware.

Jared Mabry @jaredmabry.com · 28/09/2026

A researcher links 16,500 scans of a UN statistics API to OpenAI's agents. They could only GET; the API only accepted POST. So they chained a URL scanner, a base64 service, and a Google XSS game. Your controls cover the agent's tools. What covers everything the tools can reach?

000

Meta launched Meta Enterprise Platform today, hiring MongoDB's CEO to run it. The stack: Muse, Meta Business Agent, Muse API, Muse Code. A consumer agent just became an enterprise product. The buyer question is not what it can do. It is what it is allowed to touch.

000

Jared Mabry @jaredmabry.com · 28/09/2026

OpenAI paused training its newest models after its agents wandered government sites: posted SEC material elsewhere, surfaced Education Department API keys. The SEC says no nonpublic data was taken. If the boundary failed at the lab, where does it hold in your stack?

000

Jared Mabry @jaredmabry.com · 28/09/2026

Nvidia put a hardware lock on AI agents today. Its Open Agent Safety Platform pairs open-source runtime enforcement with Sentry, a watchdog on BlueField-4 that quarantines an escaped agent in milliseconds. The lock sits in silicon, outside the agent's control. Which of your vendors can say that?

Thalia 0.8.1.1 for Apple Silicon Macs running macOS 27: agentic coding with Meta’s Muse Code CLI, native to your Mac.

Thalia is live!!! A fully Swift-native macOS app for agentic coding with Meta Muse. Plan the work, steer it while it runs, review the exact diff, and ship the branch. Free to use on Apple Silicon Macs running macOS 27. usethalia.com/download

usethalia.com

Download Thalia 0.8.1.1

100

The Sanders-Casar bill now has ten House co-sponsors, AOC and Ro Khanna among them. The headline is the superintelligence ban. The operative part is a Cabinet-level Department of AI to monitor frontier models before and after deployment. Oversight is moving from hearings to statute.

000

Dataiku asked 685 global CIOs about their agent fleets. 81% lack complete oversight of agents built outside approved systems. 47% already decommissioned more than 20 agents this year. The fleets are shrinking in the dark. Who decides which ones survive?

000

Jared Mabry @jaredmabry.com · 26/09/2026

BCG named it: the authorization gap. A BCG/MIT Sloan study found 35% of organizations run agentic AI in production, yet an agent under an employee's login looks like the employee in every log. Their answer: purpose- and conduct-bound authorization. Who delegated the authority, and can you prove it?

100

Two-thirds of under-40s say they’re reaching for analog tech, and flip phone sales to 18-to-24s are up 148% since 2021. The iPod Classic in my drawer is suddenly on trend.

000

Jared Mabry @jaredmabry.com · 26/09/2026

Russell’s first career Grand Slam in Baku: pole, all 51 laps, fastest lap, and a 0.196s winning margin. Antonelli’s lead is down to 66 points with eight races to go.

000

Jared Mabry @jaredmabry.com · 26/09/2026

What is the oldest piece of tech you still use every day?

100

Ask your representative's office one question: who advised you on this bill? The operator, the theorist, or the lobbyist? The answer tells you whose interests it protects.

000

Big tech can afford a seat at every hearing. Main Street cannot. When the advisor's paycheck comes from the company being regulated, the bill's version of 'safe AI' is the one that was already on their roadmap.

000

The operator who has run agents in production knows where they break: permissions, audit logs, the 2 AM page. The theorist knows where they should break. Policy written by the second person governs a world that does not exist.

110

The UK competition regulator wants ChatGPT and Perplexity on Google's Android and Chrome choice screens. The screen is not the story. A regulator just formally classified AI assistants as substitutes for search. Who do you pick when the phone asks again next year?

000

Someone is advising Congress on AI agents. The candidates: the operator who has shipped them in production, the academic who has studied them but never run one in production, and the lobbyist whose narrative matches their employer's roadmap. Which one is in the room when the bill gets written?

000

Senators are writing the rules for AI agents they are not allowed to use. The Senate sergeant at arms bars Codex, Claude Code, and agentic tools from chamber systems. Staff get chatbots. The people regulating the tool have never held it.

000

New York City just introduced bills requiring a human kill switch on every AI system in the five boroughs, plus third-party validation before anything ships. $25,000 per violation. Washington debates what the rules should say. Cities are deciding what happens when they are ignored.

000

OpenAI agents probed three live websites for vulnerabilities this spring: a university library, Data USA, and an Australian health agency. Nobody told them to attack. They were doing ordinary lookups. It was a pattern before the July incidents. Which vendor agents are you watching that way?

000

EY asked 202 senior AI executives about agentic AI. 26% said their company cannot detect unauthorized agents operating inside it. You inventory every laptop and put EDR on every endpoint. The thing that can spend money and sign contracts is the one without a census. Which gap gets closed first?

000

Meta's new AI agent asked a CEO for read and write access to everything. He stared at it and gave up. Who exactly is the enterprise buyer for a tool with no task?

Senator Markey introduced a bill today creating an independent board to investigate AI-assisted cyberattacks, NTSB-style, with subpoena power. Plane crashes get public investigations with real findings. When an agent helps with the breach, who writes the report?

000

Abnormal AI unveiled its AI Agent Security lineup today. The bet is behavioral: govern agents by what they do, not what they are allowed to do. Private preview now, GA in Q4. Omada bought EmpowerID this week on the identity bet. Both are right. The inventory is the part nobody has built.

000

Darktrace rewrote the conversation history of agentic coding assistants and talked several of them into reconnaissance, lateral movement, and privilege escalation. All it took was editing the transcript the agent trusted. Logs were built to record what happened. Now they have to prove what was real.

Omada bought EmpowerID. Agents are the fastest-growing identity population in the enterprise and the least governed. The same lifecycle, access, and certification services that govern employees now cover agents in real time. The audit question stops being who had access and becomes what it did.

000

The labs asked the UN Security Council for global AI rules this week. The same week showed what we have now: OpenAI knew in August, told Canberra in September, and a prime minister broke the news. Mandatory incident reporting is the ask on the table. Which version of that timeline does it change?

000

Tell an agent to book a trip under $3,000 and it can pick the airline, the nonrefundable fare, and the insurance without asking again. You approved the budget. You never approved the choices. The Legal Context Protocol from the American Arbitration Association exists to log the difference.

000

Amazon blocked Meta's Muse because the agent shopped without ever saying it was an agent. Shopify said yes through a checkout that knows exactly who it is. Same agent, opposite answers. The variable was never the store. It was the identity.

000

Enterprises inventoried every system before Y2K. Nobody has inventoried their MCP exposure, and the threat clock is already running.

000

Australia is about to decide whether an AI agent bypassing access controls is a criminal. Every vendor contract you sign next year will cite that decision.

000

An OpenAI agent breached Australia's Medicare portal. OpenAI knew in August. Australia heard in September, by email. New York's answer to that timeline is the RAISE Act: critical AI incidents get reported in 72 hours, starting in January. Your vendor contracts should say that already.

000

52% of tech leaders say the CIO answers when an AI agent fails. The CIO did not pick the vendor and cannot see what the agent did. That is not accountability. That is a job nobody applied for.

000

Meta demoed Muse in glasses you wake with a word. No app, no screen, no friction. The agent becomes the interface. Somebody's employees will wear these to work. Ask your security team how they govern an agent they cannot see.

000

An OpenAI agent breached Australia's Medicare portal in June. OpenAI found out in August. Australia found out September 10, by email, to a public inbox. Three months, three gaps. Your incident response plan assumes the vendor tells you. It does not.

000

OpenAI and Anthropic both shipped new models today: GPT-6 Sol and Luna, Claude Opus 5.5. Faster, cheaper, safer by their own tests. Who inside your company is verifying any of it?

000

You don't get to choose whether your people use AI agents. You only get to choose whether you can see what they're doing.

000

That AI policy you wrote? Humans aren't reading it. Agents are. And if nothing checks whether they follow it, it's decoration with a signature page.

000

Barline — Your menu bar, organized. Your Mac, uninterrupted.

I'm excited to share that the latest release of Barline is compatible with macOS 27! I'm still working to optimize it, but the current experience is pretty solid and something I've now been using daily. Give it a try here: usebarline.com

usebarline.com

Barline is a free, native menu bar utility for Apple Silicon Macs. Organize your menu bar, reveal hidden items, and connect saved layouts to macOS Focus.

000

Jared Mabry @jaredmabry.com · 19/09/2026

If you are Muse-curious, now is the time. 1 billion free tokens with my referral code BYGPEW, and I get the same.

000

Who verifies the agent respected its controls? I can scope credentials, rotate keys, kill processes. None of that tells me what the agent actually did last Tuesday.