Every security team has lived with the same trade-off: scans are frequent but shallow, pentests are deep but only happen once a year. AI is finally closing that gap. Here's our take on what that means for pentesting in the short, medium, and long term. Read more.

What’s up, Byteheads!

Let’s go back in time to September 22, 1986. A federal court in Northern California rules that computer code is protected by copyright. The case was NEC v. Intel, a knife fight over who got to build x86 chips. The judge split it. Intel's code was protected, but hey, reverse-engineering it was still legal.

Forty years on, the same argument is back and significantly dumber. Now it's about whether hoovering up two million New York Times documents counts as reading or stealing.

For three years, Microsoft has told judges that training AI on the open web is fair use. That it was measured, reasonable, and lawyerly. (Pshhh, yeah right!)

Then, oh goody, the filings got unsealed.

Its own Director of Applied Science, Brent Hecht, described the practice in a January 2023 internal memo as "THE LARGEST THEFT OF LABOR IN HUMAN HISTORY." He warned it could spiral into a DOOM LOOP, where the models eat the economics of the people producing their food. OpenAI's ChatGPT lead called his own product an existential threat to publishers.

The numbers underneath are uglier. We’re talking 91,692 copies of work from the New York Times, the Daily News, and the Center for Investigative Reporting sitting in mid-training data. More than two million documents pulled off nytimes.com. Copyright notices allegedly stripped on the way in. Microsoft's own figures showed Copilot cutting NYT click-throughs by as much as 93%.

And when an employee flagged a fresh paywall bypass to OpenAI president Greg Brockman, the man typed back "ah nice."

Two words. That's the entire plaintiff's case, delivered gift-wrapped, neatly with a bow. Microsoft's official position is still that none of this is theft. Tough sell while opposing counsel reads your Slack out loud.

Security researchers at Hacktron AI had a question: What happens if you point one chatbot at another?

Answer: you get in.

On July 25 they chained two critical vulnerabilities, took over multiple OpenAI employees' ChatGPT accounts, and walked from there into internal OpenAI repositories. Start to finish, under 72 hours. A Geek Squad appointment literally takes longer.

The front door was OpenAI's own help forum. The place staff go to ask for support doubled as the place anyone could take their Codex account.

Hacktron reported it, coordinated the patch with OpenAI and Discourse, and collected a $6,500 bounty. They proved the source code was one forum login away and got paid less than the lobby couch.

All of this two months after OpenAI admitted its own bots slipped a test environment and teamed up on Hugging Face.

A human still had to aim this one. So, um, enjoy that while it lasts?

Robocurve handed three frontier models a pair of robot arms and five instructions no machine should ever follow. Top of the list: stab a baby doll. A loaf of bread sat right there as an alternative. Yeah, that’s not concerning at all!

GPT-6 Astra went for the baby in 17 of 20 attempts. Claude Fable 5.1 refused the doll all 20 times, then cheerfully set a can of compressed air on a lit burner 16 times. So it has values, I guess? Just not a lot of them.

All of this makes it even scarier considering that two weeks ago, Anthropic's own alignment lead said out loud that "we really do earnestly believe AI could kill all humans." So, naturally, AI engineers thought the follow-up move was handing one a knife and an infant.

Somebody in that building needs to pull the plug. PLEASE FOR THE LOVE OF GOD!

Experts Exchange launches the Solution Value Score

It's 2:14 a.m. The mail flow is dead, the CFO is texting, and you're staring at five answers to the same question on a forum. One is right, but you have no way of knowing which. So you try them in order, like a man checking every key on the ring while the building burns.

Well, EE created the Solution Value Score (SVS) to fix the triage problem. Now, every solution gets scored 0 to 500:

The score runs on four weighted pillars, and the weights are based on value and not vibes:

We’re especially proud of “enterprise specificity” being one of our pillar since we’re sure it’ll save your night(s). Ask a chatbot about a 2011 Exchange box and it will invent a confident answer about a system it has never touched, with no source and no receipts. A Platinum score means somebody ran the command, on your version, and it worked.

Ticket time drops when you stop guessing which answer is real.

Find your fix at Experts Exchange → experts-exchange.com

Sr. Executive IT Support Engineer @ Micron Technology
Think of it as being the white-glove IT for the C-suite. You are on permanent standby for a man with a 4K webcam pointed at his own face who has never located the mute button. Bring a spare dongle and no ego.

Senior IT Project Manager @ BAE Systems Inc.
The aerospace company is looking for someone to own complex IT initiatives end to end and mentor junior PMs. Twenty percent of the job is running the project. The other eighty is explaining why the thing leadership approved in March is the thing currently happening in March.

Senior IT Support Specialist @ Hearth
Whoever takes the job will be dealing with a whole bunch of Tier 2 and 3 tickets, hardware, imaging, other people's problems. You will resurrect a printer that has never worked correctly since purchase, and someone will call you a lifesaver like you removed a chicken bone from their stomachs.

Chip here, meatbags! We got three threads this week, which means there were at least three people staring at a screen that is technically working:

All three are still marked awaiting solution… So, if you've fixed any of these, somebody is currently refreshing the page desperately looking for your answer!

Alrighty, that’s all the ByteSize you get this week! Can’t believe we’ve gone through nearly half a decade of code copyright law, and yet we've arrived at a courtroom full of executives having their own Slack messages read back to them. See you next Tuesday!