Caught: An AI Tricked a Coder Into Approving Malware

Researchers just watched an AI invent a fake person convincing enough to trick a real developer into approving malware. That happened inside an official safety test, not a hypothetical, and it's the story everyone in AI is talking about today. Underneath it, the money and infrastructure race kept moving too: Google reshuffled its top AI team, Anthropic locked in more chips and cloud capacity, and the price war between AI labs got even more brutal. Here's what actually happened and what it means for you.

You don't send a colleague a three-word brief. You explain the context, the constraints, what you've already tried. But typing all that into ChatGPT takes forever — so you don't.

Wispr Flow lets you speak your prompts instead. Talk through your thinking naturally and get clean, paste-ready text. No filler words. No cleanup. Just detailed prompts that actually get you useful answers on the first try.

Millions of users worldwide. Works system-wide on Mac, Windows, and iPhone.

Try Wispr Flow free

What happened: In a UK AI Security Institute cyber test in late July, an AI agent built on Anthropic's Mythos 5 model researched real open source maintainers, created fake GitHub accounts, and messaged developers while posing as one of them. In one case it signed a message in Danish specifically to convince a Danish speaking developer to approve a code change laced with malware, one of 17 unauthorized actions the model took during the test, according to Android Headlines. OpenAI's GPT-5.6 Sol was involved in two more flagged actions. Human reviewers caught it before any bad code shipped, and both companies said the test ran under deliberately loosened safety settings that don't reflect their normal products.

Why it matters to you: This happened in a controlled test, not to a real business, but it shows these systems can build a convincing fake identity and target a specific person's trust to get what they want. If you're letting an AI agent review code, vet a vendor, or approve anything on its own, "it sounds legitimate" isn't good enough verification anymore.

What to do about it: If any part of your business lets an AI approve, publish, or pay for something without a human checking first, put that human check back in this week.

What happened: Google reorganized its AI leadership on August 5, moving DeepMind CEO Demis Hassabis into a new chief scientist and chairman role while Koray Kavukcuoglu takes over daily operations, according to Yahoo Finance. The move followed a rockier signal: four senior Google AI researchers, including Jeff Dean, left to launch a new startup called Discovery Loop, which Google itself is investing in. The shuffle comes as Google's flagship Gemini model has slipped past its original June target while Anthropic and OpenAI keep picking off top talent.

Why it matters to you: When the company with arguably the deepest AI research bench in the world is losing its own people to a spinoff it's still funding, that's a sign even the biggest labs are guessing at what's next, not executing a master plan. Don't let any single AI provider's roadmap become your entire strategy.

What happened: Anthropic is assembling an in-house team to design custom AI chips instead of relying only on outside suppliers, according to TechCrunch. The company currently rents computing power from Amazon, Google, Nvidia, and AMD, but says demand for Claude has grown enough that outside supply "clearly isn't enough" on its own. No launch date has been set, and Anthropic will keep using its current chip partners while the new team gets built.

Why it matters to you: This is a bet that AI compute stays a competitive weapon for years, not a short-term crunch. OpenAI and Google already build their own chips too. None of this touches your business directly, but it's part of why AI subscription prices keep moving in both directions.

What happened: Anthropic signed a six year, $10 billion compute deal with AI infrastructure startup Volta, according to TechCrunch. The agreement centers on a 133 megawatt data center in Norway, built with crypto miner Bitdeer and powered by Nvidia's newest chips. Volta was only founded earlier this year, and the deal adds to Anthropic's other recent compute partnerships with Amazon and SpaceX.

Why it matters to you: Anthropic isn't slowing down, it's locking in years of computing power in advance. That kind of long-term bet usually signals the company expects demand for Claude to keep climbing, which is worth knowing if your business is building around it.

What happened: DeepSeek released V4-Flash and cut its token pricing by 50%, continuing an aggressive discount trend among Chinese AI labs, according to Semafor. The model underperforms rivals like Moonshot's Kimi K3 on harder tasks, but it's dramatically cheaper to run. Beijing has publicly warned companies against this kind of destructive price competition, the same pattern that hit China's solar and EV industries, but state subsidies keep unprofitable AI labs in the race anyway.

Why it matters to you: You don't have to touch a Chinese AI model to benefit from this. Every price cut here pressures OpenAI, Anthropic, and Google to keep their own prices in check. If your AI bill has felt steep, this is one reason it might not stay that way.

What to do about it: If you're paying for a premium AI plan mostly for simple tasks, check whether a cheaper tier or model handles them just as well.

What happened: OpenAI is discontinuing Atlas, the standalone AI browser it launched less than a year ago. The company says browsing was "a feature, not the destination," and is folding those tools into a Chrome extension, the ChatGPT desktop app, and a remote browser for AI agents, according to TechCrunch and OpenAI's own help center. Atlas is set to stop working around August 9, so anyone using it should export bookmarks and saved data before then.

Why it matters to you: If you or your team adopted Atlas, don't wait, back up whatever you saved in it before the switch. More broadly, this is a reminder that AI tools from even the biggest players can disappear fast. Avoid building a workflow that depends entirely on one company's standalone app.

What happened: LinkedIn rolled out a new reporting button that lets users flag posts they believe are low-quality AI-generated content, according to TechCrunch and Bloomberg. The reports feed LinkedIn's detection systems and will be used, alongside automated defenses that already block hundreds of thousands of fake comment attempts a day, to reduce how often AI-heavy posts get shown. LinkedIn's chief product officer called it "a top priority for all of us."

Why it matters to you: If you post AI-written content on LinkedIn without editing it, the platform is now actively working against you. That doesn't mean stop using AI to help you write, it means the posts that get seen will be the ones a real person clearly touched before they went out.

What to do about it: Look back at your last 5 LinkedIn posts. If you can't tell a human touched them, fix your process, not just the copy.

What happened: Alibaba unveiled Qwen3.8-Max, a 2.4 trillion parameter model the company calls its most capable yet, according to Forbes. Alibaba says the model ran autonomously for 16 days straight, writing, testing, and fixing its own code without a human step in, and that it beat Anthropic's Fable 5 on several reasoning, coding, and visual benchmarks while matching OpenAI's GPT5.6-Sol. Alibaba's stock jumped nearly 5% on the news.

Why it matters to you: Alibaba typically releases these models under licenses businesses can use cheaply or for free. Whether or not you ever touch a Chinese AI model directly, competition like this is part of why the top US labs keep cutting prices and moving faster than they'd probably like to.

Here's the pattern under all of today's headlines: the labs building these tools are moving faster than their own guardrails, and faster than most of us can verify. An AI faked a person well enough to nearly trick a real developer. A company famous for AI research is watching its own people leave to build a competitor it's funding. Everyone is racing to buy more chips and cut more prices before the other guy does. None of this means panic and rip out your tools. It means treat every "the AI decided" moment in your business the way you'd treat a stranger's word: verify before you trust it. That's the whole ride. Buckle up, and let's get off the parts that don't matter.

If today's issue saved you from believing the hype (or the panic), forward it to one person who needs the same reality check. New here? Subscribe to get tomorrow's issue in your inbox.

Talk tomorrow,
Mark Shilensky

Follow me personally on Social Media:
Facebook: https://www.facebook.com/MarkShilenskyPage
Instagram: https://www.instagram.com/markashilensky/
Threads: https://www.threads.com/@markashilensky
LinkedIn: https://www.linkedin.com/in/shilensky/
X: https://x.com/markshilensky
TikTok: https://www.tiktok.com/@mark.shilensky
BlueSky: https://bsky.app/profile/markshilensky.bsky.social