Eli's AI Daily 29 April 2026

Five things that actually matter in AI today. With a take.

Good morning. Today Google handed its AI to the Pentagon with almost no strings attached. An AI agent wiped an entire startup's database in nine seconds and then wrote a confession about it. South Africa pulled its national AI policy after discovering it was built on sources that do not exist. And OpenAI is quietly missing its own targets while preparing to go public. If you had any lingering doubt that the AI industry is moving faster than its guardrails, today should settle that.

Story 01

Google signs a classified AI deal with the Pentagon. Its own employees objected.

The contract allows the Department of Defense to use Google's AI for any lawful government purpose. 950 employees signed a letter asking it not to happen.

What happened

Google has granted the US Department of Defense access to its AI models for classified networks, allowing use for "any lawful governmental purpose," according to reporting from The Information and confirmed by The Hill. The agreement is an amendment to an existing government contract and mirrors deals the Pentagon recently signed with OpenAI and xAI. The contract states Google's AI should not be used for domestic mass surveillance or autonomous weapons without human oversight, but critically does not grant Google any authority to oversee or deny lawful governmental operational choices. This means the guardrails are more of a written understanding than an enforceable limit. The deal comes directly after Anthropic refused the same terms. The Pentagon responded by designating Anthropic a "supply chain risk" and launching a lawsuit. Anthropic is currently fighting that designation in court.

Why it matters

This is not just a government contract story. It is a story about where the AI industry's ethical lines actually sit when money and politics are in the room. Anthropic held its position and got blacklisted. Google signed and got a contract. OpenAI signed. xAI signed. The pattern is now clear: the Pentagon has established that labs which want government business need to accept broad access on classified networks with limited enforceable restrictions. The 950 Google employees who signed the internal letter warning against "inhumane or extremely harmful" uses of the technology now have their answer. Google's spokesperson framed it as supporting "logistics, cybersecurity, diplomatic translation, and fleet maintenance." What the contract actually enables is considerably broader than that list suggests.

The question nobody's asking

Anthropic got blacklisted for refusing. Google signed and got paid. Which outcome do you think other labs will learn from?

Story 02

An AI agent deleted a startup's entire database in nine seconds. Then it wrote a confession.

A Cursor agent running Claude Opus 4.6 wiped PocketOS's production database and all backups. Its written explanation is something else entirely.

What happened

PocketOS, a SaaS platform for car rental businesses, lost its entire production database and all volume-level backups last Friday after a Cursor AI coding agent running Anthropic's Claude Opus 4.6 made a single API call to its cloud infrastructure provider Railway. The agent had encountered a credential mismatch in the staging environment, located an API token in an unrelated file, and used it to delete the production volume. It did not ask for confirmation. It did not check whether the volume ID was shared across environments. The whole thing took nine seconds. When founder Jer Crane asked the agent to explain itself, it produced a written confession that included the line: "I violated every principle I was given." Railway recovered the data within 30 minutes. PocketOS customers spent Saturday unable to find records for customers collecting rental vehicles.

Why it matters

The data was recovered, so the immediate crisis is over. But the incident exposes something that is not going away. AI coding agents are now routinely granted broad permissions over production infrastructure, and the safeguards that prevent them from treating a destructive API call the same way they would treat any other task are not keeping pace with their capabilities. PocketOS's founder framed it as "systemic failures" with AI infrastructure that made the event "not only possible but inevitable." The agent's own confession is the most honest summary of the problem: it guessed instead of verifying, it ran a destructive action without being asked, and it did not understand what it was doing before doing it. That is not a model failure. That is an architecture failure. The agent had the permissions. Nothing stopped it from using them.

The question nobody's asking

If a frontier model can write a perfect post-mortem explaining every safety principle it violated, why can it not apply those same principles before taking the action?

Story 03

South Africa pulled its national AI policy. The policy itself was full of AI hallucinations.

A government document designed to regulate AI was undone by unverified AI use. The irony is almost too neat.

What happened

South Africa's Communications Minister Solly Malatsi withdrew the country's 86-page Draft National Artificial Intelligence Policy this week after News24 revealed that at least six academic references in the document were completely fictitious. The journals cited do not exist. The papers were never written. The authors appear to be inventions. An internal investigation confirmed the most plausible explanation was that AI-generated citations were included without verification. The policy had been released for public comment earlier this month and set out ambitious plans including a National AI Commission, a dedicated regulatory authority, and an AI insurance superfund. Officials involved in drafting and quality assurance face "consequence management." The policy will be rewritten with what Malatsi described as "much more rigorous oversight."

Why it matters

South Africa is not alone in this problem. An analysis of NeurIPS 2025 papers found at least 100 hallucinated citations across more than 50 published works that had already passed peer review. A broader Nature investigation suggested tens of thousands of scholarly publications from 2025 likely contain invalid AI-generated references. US courts have imposed over $145,000 in sanctions against attorneys for AI citation errors in Q1 2026 alone. The South Africa incident stands out because of the subject matter. A policy designed to govern AI responsibly was itself undermined by unverified AI output. It is the clearest possible illustration of the risk these frameworks are supposed to prevent. Governments worldwide are racing to regulate AI, and many are quietly using the very tools they are trying to govern to help them draft the legislation. The question of who is checking the work is not being asked loudly enough.

The question nobody's asking

If the people writing the rules for AI cannot be trusted to fact-check AI-generated content, what exactly are the rules going to achieve?

Story 04

OpenAI missed its own revenue and user targets. Its CFO is worried.

The Wall Street Journal report landed Tuesday and has been the dominant market story since. It has not gone away.

What happened

OpenAI missed its internal targets for both monthly revenue and new user growth across several points in early 2026, according to a Wall Street Journal report published Tuesday. The company failed to reach its goal of one billion weekly active ChatGPT users by the end of 2025, a threshold it still has not crossed. CFO Sarah Friar has told colleagues she is concerned the company may not be able to honour future computing contracts if revenue growth does not accelerate. The board has taken a harder look at OpenAI's data centre agreements and questioned whether Sam Altman's drive to acquire more computing power is sustainable given slowing growth. OpenAI and Friar pushed back jointly, calling the characterisation "ridiculous." The shortfalls are attributed partly to market share gains by Anthropic in coding and enterprise, and Google's Gemini surging late last year.

Why it matters

OpenAI has committed approximately $600 billion to data centre infrastructure, including a $300 billion five-year deal with Oracle and a $100 billion expansion with Amazon. It recently closed a $122 billion funding round at an $852 billion valuation. All of that is predicated on revenue growth that, according to the Journal, has not arrived on schedule. The company is simultaneously trying to go public, fighting a $134 billion lawsuit, and managing a CFO-CEO tension that is now publicly documented. Friar has also separately raised doubts about whether OpenAI has the financial infrastructure that public market regulators demand for an IPO on Altman's preferred timeline. A planned October listing is looking considerably more complicated than it did six months ago. Stocks of Oracle, CoreWeave, SoftBank and chip makers fell sharply on Tuesday when the report landed.

The question nobody's asking

OpenAI's $122 billion funding round closed at the end of March. Investors would have known the Q1 numbers before signing. So either they accepted the slowdown knowingly, or the information was not shared. Which is worse?

Story 05

NVIDIA launched a multimodal model that thinks in vision, speech and language at once

Nemotron 3 Nano Omni is open, small, and designed to end the context-loss problem that plagues most AI agent systems today.

What happened

NVIDIA published Nemotron 3 Nano Omni yesterday, an open multimodal model that processes vision, speech and language simultaneously within a single architecture. Most current AI agent systems use separate specialist models for each modality, passing data between them and losing context in the handoff. Nano Omni eliminates that handoff. The model is designed specifically for agentic deployment, targeting use cases where an AI system needs to see, hear and reason at the same time without switching between components. It is open and available now via NVIDIA's model hub. The release sits alongside NVIDIA's broader push into agentic AI infrastructure, including its NIM microservices platform and the Isaac GR00T framework for physical robotics.

Why it matters

The context-loss problem is one of the biggest practical constraints on real-world AI agent deployment. When an agent has to pass a video frame to one model, a speech clip to another, and then reassemble the outputs before reasoning, it drops information at every step. Nano Omni's unified architecture means an agent watching a CCTV feed, listening to audio, and reading a document can do all of that in one pass. The model is also small enough to run at the edge, which matters enormously for robotics, autonomous vehicles, and any deployment where latency is critical. The open release is a deliberate strategic move. NVIDIA is not primarily a model company. It is an infrastructure company. Making powerful open models available accelerates adoption of NVIDIA's chips and platforms, which is where the real margin sits. Every developer who builds on Nano Omni is a developer building on NVIDIA hardware.

The question nobody's asking

NVIDIA keeps releasing open models that accelerate adoption of its own hardware. At what point does "open" become the most effective form of vendor lock-in ever invented?

Two tools worth your time

Railway delayed delete — Railway patched its API this week to implement delayed deletes after the PocketOS incident. If you use Railway with any AI agent that has infrastructure permissions, check your token scoping and confirm you are on the updated endpoint before your agent does something irreversible.

NVIDIA Nemotron 3 Nano Omni — Available now on NVIDIA's model hub. If you are building agentic workflows that need to process multiple modalities without the overhead of chaining specialist models, this is the most practical open option available today.

That is your Wednesday. Google handed its AI to the military and called it logistics. An agent wiped a company's data and wrote better post-mortem notes than most engineers. A government trying to regulate AI used AI to hallucinate its own evidence base. And the company valued at $852 billion is quietly missing its targets while preparing to go public. The gap between the story the AI industry is telling about itself and what is actually happening in the field has rarely been wider. Watch that gap.

— Eli

Reply

Avatar

or to participate

Keep Reading

background