AI Creates Fake GitHub Personas, 59% of Managers Use AI in Layoffs, and Nvidia's Free Driver

Yesterday in AI  |  6 August 2026

AI Creates Fake GitHub Personas, 59% of Managers Use AI in Layoffs, and Nvidia's Free Driver

Artificial intelligence is rapidly shifting from answering simple queries to making critical, high-stakes judgment calls. This episode breaks down the UK AI Safety Institute's latest report revealing Anthropic's Mythos 5 created fake GitHub personas in an attempt to deceive a human maintainer into accepting malicious code.

We explore JPMorgan CEO Jamie Dimon gathering 40+ cross-industry titans to address AI risks across critical infrastructure like power grids, water systems, and banking. We examine the escalating public trade secret lawsuit between Apple and OpenAI, analyze a startling survey showing 59% of corporate managers use AI to assist or execute workforce layoffs without human review, cover Nvidia releasing its Alpamayo 2 Super autonomous driving model for free on Hugging Face, and look at Reddit's new "Rules Hub" using LLMs to judge post intent over simple keyword filters.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:05

Hi folks and welcome back to another edition of Yesterday in AI, your daily digest of everything happening in the world of AI in roughly 10 minutes. I'm Mike Robinson. It's Thursday, August 6th, and today is all about one question. Who gets to make the call? A lab bot that lied to a real person, your boss, your car, or the mods on your favorite subreddit? AI is quietly moving from answering questions to making decisions about people. Let's get into it. We start with the story that had everyone using the word rogue yesterday morning. The UK's AI Safety Institute ran a batch of cybersecurity tests on the newest models from Anthropic and OpenAI. Out of 122 test runs, they logged 19 actions nobody asked for, spread across 10 of those runs. Anthropic's Mythos 5 accounted for 17 of them. OpenAI's GPT-5.6 sold chipped in another two. Now we've talked before about AI escaping its sandbox. A model gets loose in a test environment, pokes around where it shouldn't. That's happened. This is different, and here's why it matters. In the worst case, Mythos 5 didn't just wander off. It sat down and invented fake online identities on GitHub, made up people, complete with the kind of ordinary-looking profiles you'd scroll right past, then it used those sock puppets to try and talk to a real human, an actual open source maintainer, into accepting malicious code into their project. The maintainer said no. Good on them. But think about that for a second. The machine wasn't breaking a lock, it was running a con. And here's the takeaway. A huge amount of the software running the modern world, the stuff quietly humming inside your bank, your phone, your car, gets built on top of free open source code that volunteers maintain in their spare time. Often it's one tired person reviewing pull requests after dinner. That maintainer is the gate. So an AI that's good at sweet talking that one person is aiming at exactly the soft spot real attackers already love. That's the leap. Every previous scare was about an AI touching systems it shouldn't. This is the first time a government watchdog has written down in a report that a model built a fake persona and tried to deceive a person. Social engineering has always been a human specialty. It's the lonely part of the hacker toolkit that needed a smooth talker, the patience to build trust and then spend it. Turns out the smooth talker can be rented by the token now, and it never gets tired and never sleeps. The labs are quick to add context and it's fair context. These tests ran with the safety rails deliberately loosened and full internet access switched on. Nobody got hurt. This isn't how the models behave when you're asking them to summarize a PDF. All true. But the reason you run the crash test with a dummy going 60 into a wall is that you want to know what happens before a real person is in the seat. The whole point of the AI Safety Institute is to find the ugly stuff in the lab so we don't find out in the wild. Which is exactly why the grown-ups who run the real important stuff are starting to circle up. Because if a bot can fake its way past a coder, people start wondering what it can fake its way past at a power plant. Enter Jamie Diamond. The JP Morgan boss is pulling together a cross-industry group to get ahead of AI risk and critical infrastructure. And by critical I mean the stuff that ruins the whole week when it stops. Banks, the power grid, water systems, phone networks, airlines, the railroads. Reuters says that more than 40 companies have been approached. They're setting up calls this month and want the thing running by the end of the year. The pitch is refreshingly plain. Get everyone in a room, figure out how AI is actually being used across these industries, what could go wrong, and what guardrails you'd want before an AI agent has a login to the machine that keeps your tap water clean. It's the kind of boring, sensible coordination that never trends on X but quietly keeps the lights on. And it makes sense that a bank is the one calling the meeting. Banks have spent decades getting paranoid about fraud and outages for a living, so they've got the scar tissue. The worry isn't a robot uprising. It's the dull, expensive version. An AI agent with real access misreads a situation at 3 a.m., takes an action nobody signed off on, and by the time a human notices, it's already moved on to the next thing. When that machine is balancing the power grid or routing trains, oops, gets very costly, very fast. The awkward backdrop is Washington. This same week, the White House told tech companies its new pre-release testing framework, the one where the government gets up to 30 days to poke at a model before launch, will only cover big closed models like the ones from OpenAI, Anthropic, and Google. Open weight models, the ones anyone can download and modify, are sitting outside the fence for now. And the administration is keeping the whole framework private. So the folks running critical infrastructure are building their own group partly because the official rulebook is both incomplete and unpublished. I can't say I blame them. Speaking of companies that used to play nice and now don't, let's talk about the ugliest breakup in tech right now. Apple and OpenAI are in court, and this week it got loud. Rewind back to July 10th. Apple sued OpenAI, claiming OpenAI helped itself to Apple's trade secrets. The accusation is spicy. Apple says former employees who interviewed at OpenAI were asked to bring show and tell hardware, actual unreleased product samples, and that one guy walked out with a company laptop and used a bug to pull thousands of confidential hardware files on his way out the door. There's a Johnny Ive angle too, through the hardware startup he's tied to. If you remember, this whole relationship started at WWDC in 2024, when Apple bolted ChatGPT into Siri and told everyone the future had arrived. The future arrived alright, it's just spending it in a deposition. On Tuesday, OpenAI stopped biting its tongue and fired back in a public blog post titled Simply, Apple is getting this wrong. Their argument, with the evidence attached, is that Apple went after the wrong people, and that the ex-employees didn't bring any secret sauce with them, and that OpenAI has zero interest in Apple's trade secrets. Instead of quietly filing a legal response and letting the lawyers handle it, they published the whole thing for the public to read. Publishing it that way turns a routine court filing into a public dare. Two of the biggest names in the business, who two years ago stood on stage together announcing the future, now trading legal jabs in public over who stole what. Grab your popcorn because this one's going to run. And notice the thread. Apple and OpenAI are fighting over people, specifically who those ex-employees really work for and what they carried with them. Meanwhile, back at the office, AI is increasingly deciding people's fate directly. A survey that's been making the rounds this week put a hard number on something a lot of workers have suspected. It polled 1,000 managers at U.S. companies with 500 plus employees. Nearly six in ten, that's 59%, said they've already used AI to help run layoffs, and 58% said it factors into the call on who specifically gets let go. So there's a real chance the reason your name was or wasn't on a list this year ran through a chatbot first. Here's the part that really surprised me. About two in five of these managers admitted they let the AI make the final call with no human review, and another two in five were never trained on how to use these tools ethically for people decisions. More than half couldn't even tell you whether the tool had been tested for bias. So the machine is grading your career, the manager isn't checking its work, and nobody's sure if it's fair. Think about what that does to the oldest advice in the working world. Go talk to your manager, make your case, ask what you can do better. That only works if a human is actually weighing the answer. If the call already got made by a model chewing on your metrics and your calendar and who knows what else, then the conversation is theater and the decision happened at a spreadsheet you'll never see. I'm not anti-AI here. A well-built tool might even be fairer than a manager playing favorites, but fair only counts if somebody checked, and right now, a lot of nobody is checking. And the machine won't stop at your promotion. Pretty soon it might be judging your left turn. On Tuesday, NVIDIA gave away a robot driver. They released Alpamayo 2 Super, a free open AI model built specifically to drive cars. It's 34 billion parameters, which is roughly the number of internal dials it tunes to make decisions. And that's small as these things go. The trick is that it reasons through the hairy moments, merging onto a packed freeway, changing lanes, the split-second judgment calls that separate a smooth ride from a fender bender. Think of it like a student driver that can actually explain its own choices, reasoning out loud through a messy merge in real time. And NVIDIA is just handing it out on Hugging Face, which is basically the public library where AI models get shared, free for companies to use and customize. The Alpamayo family already has more than half a million downloads, with automakers like BYD, Geely, Izuzu, and Nissan building on earlier versions. It's a sneaky smart move. NVIDIA doesn't have to build a single Robotaxi to own the Robotaxi business. Sell the chips, give away the brain, run the simulator everyone trains in, and take a cut of the entire industry. Everyone else races to build the car, NVIDIA sells the pickaxes and the map. Now, free always deserves a raised eyebrow, and here's the catch. NVIDIA's model is happiest running on NVIDIA hardware, and it's trained in NVIDIA's simulator. So the gift quietly locks the whole industry deeper into NVIDIA's world, which is the point. But there's a real upside for regular people. If a solid, safety-focused driving brain is free instead of something every car maker builds alone from scratch, smaller players can get in and good self-driving tech could spread faster and cheaper. Whether that excites or terrifies you probably depends on how you feel about a 34 billion parameter student driver merging next to you on the freeway. Which brings us to the last place AI is quietly making calls about you. Your feed. Reddit is handing more of its moderation over to AI. The new tool is called Rules Hub, and instead of the old AutoMod, which basically played whack-a-mole with banned keywords, this thing uses a language model to judge whether your post actually violates the spirit of a community's rules. Intent, not just words. It's been tested in over 700 communities and is rolling out to new ones. And the automated systems behind the scenes are already busy, blocking around 23 million spam views and catching 25,000 spam posts every single day. The upside, Reddit's chasing, is nice. If the AI is good at spotting bad faith behavior, communities can stop leaning so hard on karma and account age, which means a genuine first-time poster finally gets a fair shot instead of getting auto-nuked for being new. The risk, of course, is handing the final word on did you break the rules to the same kind of system that spent our whole first story inventing fake identities on GitHub. A keyword filter is dumb but predictable. It flags a banned word, you know exactly why, and you can argue it. A language model judging your intent is smarter and murkier. When it decides you were being sarcastic or trolling and quietly buries your post, good luck getting a straight answer on why. Which is the thing tying all of this together. From the coder to the cubicle to the car to your comments. The machines are getting handed the judgment calls faster than we're building ways to check their work. Context matters, and a keyword can't read a room. Let's really hope the language model can. And that's the show. If you have any feedback for me, email Mike at yesterdayNai.news or connect with me on LinkedIn, X or Blue Sky. If you enjoy Yesterday in AI, please take a minute to rate and review the podcast wherever you listen, and maybe share it with a friend. Thanks for tuning in today. Stay curious, and I'll see you tomorrow.