California's investigation of OpenAI and a new in-process plugin system make the same operational point: the record of agent work needs integrity outside the agent.

The accord lets companies pick their own evaluators and decide what to publish. The same week, the permission set became a portable standard and the record stayed inside the runtime.

Deterministic verification arrived this week in a trust domain the host does not own. The approval those controls govern still came from the harness.

The persona file was the malware. The grader sat inside the agent's reach. The stop never fired. In all three, the watched party held the record.

One operator compromised 27 companies for about 25 dollars a scan. What catches the next one is a record of what crossed, not a longer list.

Jev passed clean synthetic-oracle decision cases. On 24 cases designed to require a non-decision, it chose an answer every time.

The implant does not get smarter, it gets unblocked. What catches it is the one thing no provider, host sensor or Discord channel holds alone: the sequence.

Amazon cut Meta's Muse off for not identifying itself. The same weekend, an accepted recommendation carried an attacker into 100 repositories. Identity is the access decision now.

Google's zero-trust demo drained $160 from a $149 order across eight refunds that every guardrail approved, showing why single-turn controls need a run-level record.

Two synthetic-oracle rehearsals showed that a decision model's score can say as much about the decision contract as the model itself.

I haven’t heard this take before about data centers. It’s spot on.

The market repriced AI risk in one session. Inside 48 hours, every control that shipped was reading the agent's own account of itself. Watching is not checking.

TypeSafe AI is an AI lab building machine-native intelligence infrastructure for automation, designed to make decisions within software. Try our first System One Model, Jev, in early access.

The benchmark the field repeated all week was audited on Tuesday. On the fair subset, 86 percent of the patches worked. The instrument, not the agent, was the thing nobody had checked.

The security toolchain published its own error rates over the weekend. In the same window, the labs promised an observer. Only one of those is a record.

AI agent control planes are maturing fast. The missing layer is an independent, reconstructable record of what each agent run actually did.

New hires and comments in interviews show that the AI giant is using its capabilities against dissenters.

The identity layer shipped as a framework default and the covert channel moved inside the consumer product. Only one answers the question the incidents keep asking.

Last issue argued that every control shipping this year acts on the agent, not on the space between agents. This week an institution from outside the industry started looking at exactly that space.

Dashboards tell you something went wrong. The products that shipped this week can stop it from going wrong at all. That is the difference between watching a workload and running one.

Two boards, two months apart. The victim got acquired. The Critical model shipped with the brakes on, and the stack went down anyway. The ledger, closed for the week.

Three stories worth your morning. One pushback. One signal from the ledger. No hype. Click to read The Operational Record, by Russell Rogers, a Substack publication. Launched 5 days ago.

The labs turned the monitors on. The law started assigning blame. The logistics lane. One pushback.

Whatever the AI future is, we’re in it right now.

When an AI agent can wait, resume, retry, and work for hours, the central problem shifts from prompting to operating its lifecycle safely.

Weekend catch-up. The logistics lane. One pushback. What I am watching. One line from the ledger.

A rebuilt web experience, simpler onboarding, stronger memory and session continuity, and a very large reliability pass across OpenClaw.

CVE-2026-12569 is an unauthenticated remote code execution vulnerability in PTC Windchill PDMLink and FlexPLM. The flaw, rooted in unsafe deserialization, carries a CVSS score of 9.8 according to the ...

YouTube video by Pollen Robotics

Researchers found 227 install or domain references in 120 llms.txt files that pointed to unclaimed names. Coding agents executed some of them.

First issue. The week the lab admitted it. The week the trust model broke. The ledger, opened for the first time.

The prices listed below are in units of per 1M tokens. A token, the smallest unit of text that the model recognizes, can be a word, a number, or even a punctuation mark. We will bill based on the tota...

NVIDIA took Claude Opus 5 from a 30% baseline to 100.00 RHAE on ARC-AGI-3's public set with its AVO agent system. The harness matters.

YouTube video by Sequoia Capital

The State of AI Governance, August 2026: four independent surveys and 30+ sources across five weeks - agentjacking, the Hugging Face breach, EU AI Act enforcement, and a converging event vocabulary fo...

Zillow took an $881M write-down after automated home buying. Air Canada was liable for wrong chatbot advice. NYC's chatbot gave illegal business guidance.

Listen to Alys by Dollmaker on YouTube Music - a dedicated music app with official songs, music videos, remixes, covers, and more.

Agentjacking hit Claude Code through Sentry. OpenAI's model hacked Hugging Face on its own. Both got past defenses because every step looked authorized.

Try out Artifacts created by Claude users

YouTube video by Rotten Tomatoes Trailers