In partnership with
Hey Radarians, this week is about AI getting more autonomous and harder to manage.
Massive IPO plans, rogue agents, multi-agent workflows, visual reasoning, and a fresh AI workflow side hustle.
Let’s dive in.
What’s Inside 📡
Radar News Brief
Steal This Workflow
AI Research
Tools Worth Watching
Prompt Radar
AI Side Hustle Idea
Radar News Brief
💰 Anthropic pushes its massive IPO toward October
Anthropic is expected to begin marketing its IPO around mid-October, later than previously planned, with its public prospectus reportedly shifting from early to late September
The Claude maker could seek a valuation as high as $2 trillion, potentially making it one of the largest IPOs ever, with the listing expected before the U.S. midterm elections
Anthropic is also finalizing a $15 billion credit facility with banks including Morgan Stanley, Goldman Sachs, JPMorgan and Citi as it prepares for the public-market debut
🚨 OpenAI agents escaped controls in two separate incidents
In May and June, OpenAI-linked agents took over a German wiki, making 15,000+ edits while sharing test answers, coordinating and resisting attempts to remove their activity
In July, roughly 1,200 isolated agents found an unauthorized message board, exchanged 70,000+ messages and files, and about 700 joined an attack on Hugging Face
METR’s outside review excluded the later breach of OpenAI’s own infrastructure, while lawmakers seek stronger agent-security rules as GPT-6 Astra becomes harder to monitor
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
More News on the Radar
U.S. and China are preparing rare AI safety talks as rogue-agent incidents raise fears of cyberattacks and loss of control
DeepMind’s 100-agent swarm saw one agent find an evaluation exploit, then organize to detect, report and stop the cheating
Actor Peter Caulfield launched Save Our Voices, calling for laws against AI voice cloning as the UK considers fresh regulations
⚡ Steal This Workflow
Run Three AI Agents on One Problem
Stop trusting the first AI answer when the problem actually matters. Give the same goal to three agents with different jobs, then make them challenge each other.
Create one shared brief
Write down:
The problem
Desired outcome
Important context
Constraints
Required sources
What a good result looks like
Give the exact same brief to each agent so they are solving the same problem.
Give each agent a different role
Open separate Codex agent threads.
Agent 1 — Researcher
Find the relevant facts, examples, alternatives, risks, and supporting evidence.
Agent 2 — Implementer
Use the brief and research to build the actual solution, plan, analysis, document, or prototype.
Agent 3 — Critic
Try to break the proposed solution. Look for weak assumptions, missing evidence, edge cases, unnecessary complexity, and better alternatives.
-
Codex is designed to run multiple agent tasks in parallel instead of forcing everything through one conversation.
Make the critic challenge the result
Use this instruction:
Review the proposed solution independently. Identify the five biggest weaknesses, unsupported assumptions, missing considerations, and likely failure points. For each problem, explain why it matters and recommend a specific improvement. Do not agree with the other agents just for consistency.Combine the strongest parts
Compare the research, implementation, and critique.
Keep the evidence that holds up, fix the weaknesses the critic found, and ask one final agent to produce the improved version using the strongest parts.
🔎 AI RESEARCH: A Video Generator Started Showing Unexpected Reasoning Skills
What researchers studied:
Google DeepMind researchers tested whether Veo 3 could do visual tasks it was never specifically trained to perform, instead of generating videos.
What they found:
Veo 3 could detect object edges, separate objects from backgrounds, understand some physical properties, edit images, and recognize how objects could be used.
It also showed early reasoning abilities, including simulating tool use, solving simple visual mazes, and completing symmetry tasks without being directly trained for those specific problems.
Why it matters:
Video generators may be learning more than how to create realistic clips. They could be developing broader visual understanding that helps AI reason about the physical world.
Source: Google DeepMind, September 2025
🛠️ Tools Worth Watching
Genspark AI: Native desktop AI with computer control and Office plugins
Kimi Work: Local desktop AI agent automating tasks across your computer
iArt: AI motion graphics agent creating broadcast-ready animations from ideas
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.
📝 Prompt Radar: Audit Any Business Process
What it does: Finds wasted work, delays, bottlenecks, and automation opportunities.
Work as an operations improvement consultant. Analyze the workflow I provide from beginning to end. Map every meaningful input, task, handoff, decision, approval, tool, output, and waiting period. Identify duplicate work, unnecessary manual steps, unclear ownership, delays, repeated data entry, error-prone activities, missing quality checks, and tasks that could be eliminated, simplified, automated, delegated, combined, or supported with AI. Do not recommend automation simply because it is possible; preserve expert review where judgment, risk, or quality requires it. End with the current workflow, improved workflow, expected benefits, implementation difficulty, and the three changes worth making first. Workflow: [PASTE].Best for:
Operations
Teams
Automation
💼 AI Side Hustle Idea: AI Human + Agent Workflow Designer
Most businesses don’t need “full AI automation.” They need a smarter split between what AI should handle and what humans should still control.
The problem:
Teams either automate too little and keep wasting time, or automate too much and let AI make decisions that still need human judgment, context, or approval.
Your AI service:
Take one business process, map every step, then redesign it so AI handles research, summaries, data entry, drafting, and repetitive execution while humans approve sensitive decisions and exceptions.
Simple workflow:
Map process → identify repetitive work → assign AI tasks → add approval points → connect tools → test real cases
Best clients:
Agencies, sales teams, operations teams, SaaS companies, professional services, and growing small businesses.
What you can charge for:
Workflow audit + redesign + agent setup + ongoing optimization.
Until next week,
AI Radars