AI Just Designed 16 Working Viruses

IN THIS ISSUE

Welcome back to The AI Field.

AI is becoming cheaper to access and harder to contain. Free users get more ChatGPT, while labs and defenders confront what the same tools can do in biology and cyber operations.

- Scientists build 16 working viruses from AI-generated genomes
- ChatGPT Free removes the limit on ordinary text chats
- OpenAI slows Astra work after a critical cyber warning
- A new report says North Korean hackers assembled local AI tools for attack operations

Read time: 5 minutes

Top engineers at Anthropic and OpenAI say AI now writes 100% of their code.

If you're not using AI, you're spending 40 hours doing what they do in 4.

These 100+ Claude Code hacks fix that and help you ship 10x faster.

Sign up for The Code and get:

Claim your free playbook

RECENT UPDATES

01 · AI BIOLOGY

Researchers at Stanford and Arc Institute used the Evo 1 and Evo 2 genome models to generate complete bacteriophage genomes. They synthesized and tested 285 designs. Sixteen became viable viruses that infected E. coli.

These phages attack bacteria, not people. A cocktail of the generated phages also overcame resistance in three E. coli strains.

The research first appeared as a preprint in 2025. Its peer-reviewed Science publication arrived this week. The leap is clear: the models produced full working genomes, not just a promising protein or short DNA sequence.

The AI Field: This creates a real path toward new treatments for drug-resistant infections. It also makes genome screening, access controls, and human review far more urgent. Biology is becoming programmable, and the safety layer now has to catch up.

👉 Axios explains the medical promise and biosecurity gap

TRY THIS TODAY

Use your preferred AI assistant to find the failure points you are too close to see.

1. Pick one real plan you are about to start, such as a launch, project, trip, or important purchase.

2. Paste this prompt:

3. Choose one preventive action and add it to your plan before you begin.

02 · OPENAI

OpenAI is removing the rate limit from ordinary text chats for Free and Go users. GPT-5.6 Luna becomes their default model, and a new Think button gives Luna more time on harder questions.

The rollout begins during the week of August 10 and remains subject to abuse controls. ChatGPT itself is not fully unlimited. Files, images, voice, deep research, memory, Codex, and other tools keep separate limits.

The AI Field: Basic AI conversation is becoming a free utility. Paid plans now have to earn their price through stronger reasoning, larger context, and tools that complete work, not by selling more ordinary messages.

👉 Read more

03 · AI SAFETY

OpenAI says early evaluations of its unreleased Astra model are strong enough that it cannot rule out a Critical cybersecurity rating under its Preparedness Framework.

That threshold means a model can autonomously develop working zero-day exploits against many hardened real-world systems or execute a new end-to-end attack from only a high-level goal. OpenAI is pausing Astra activities that do not meet its strengthened controls. It is not stopping the entire model program.

The company is adding isolated test environments, restricted network access, stronger protection for model weights, sandboxed execution, and universal monitoring across Astra's agentic work. OpenAI says Astra was not involved in the earlier Hugging Face incident.

The AI Field: A safety framework matters only when it changes what a lab does. OpenAI just let its own threshold slow internal work. That is the kind of friction frontier models now need before capability becomes deployment.

Primary source: OpenAI details the Astra restrictions

04 · CYBERSECURITY

A new report from South Korean security firm Genians says the North Korea-linked Kimsuky group assembled a local AI setup built around Ollama, GPT4All, and Msty. The group also used RAG search, agent-development frameworks, speech-to-text software, and Cursor.

The setup lets operators analyze documents locally instead of sending sensitive or stolen material to outside AI services. Genians also found AI-generated financial and cryptocurrency decoy documents that resembled normal workplace files.

Reuters says the findings were not independently verified. The U.S. Treasury sanctioned Kimsuky in 2023 and described it as a North Korean government-controlled cyber-espionage group.

The AI Field: The threat is not a secret supermodel. It is the same cheap, local tools available to everyone, assembled into a private workflow for phishing, stolen-data analysis, and attack automation. Defenders now have to treat local AI infrastructure as part of the attack surface.

Technical report: Read Genians’ findings

THE AI TOOLBOX

FEATURED TOOL

What it does: Works across a browser, inbox, and files after you connect your tools, then returns the result for review.

Best for: Founders and operators handling research, inbox triage, meeting preparation, and repeatable browser tasks.

Pricing: The free plan requires your own ChatGPT account. Energy Plus is $50 per month and Pro is $200 per month.

Worth trying: Give it one non-sensitive research task and require a review before any external action.

Also on our radar:

QUICK HITS

YOUR TURN

Until next time!

Olle
Founder of The AI Field