Bug Bounty Researchers Hacked OpenAI Using Anthropic's Claude - DTH

Duration: 5:23 756 transcript words EN summary

Video summary

A three-person security team at the startup Hackron AI recently utilized Anthropic's Claude model to successfully exploit a vulnerability in OpenAI's Discourse-based staff forum, gaining unauthorized access to employee ChatGPT accounts and a GitHub repository. This incident occurred while the team was participating in OpenAI's bug bounty program, leading to a payout of $6,500 after they reported the flaw. Although OpenAI has since patched the security issues, this event highlights a growing concern that AI tools are making sophisticated hacking attacks cheaper and faster for bad actors. The transcript notes that this is merely another example of how artificial intelligence is being weaponized to bypass traditional security measures more efficiently than before. In legal and regulatory news, newly unsealed court documents reveal stark internal warnings from Microsoft and OpenAI regarding the dangers of AI scraping. Brent Heck, a director at Microsoft Applied Science, described AI data scraping as an "astonishing theft" and potentially the largest theft of labor in human history, while OpenAI's Nick Turley warned that publishers face an existential threat from chatbots trained on their news content. Separately, a German court ruled that Meta could be held liable for fraudulent investment ads on its platforms, ordering the company to remove the content, pay damages, and disclose revenue details because its algorithms actively control what users see, negating claims of lack of knowledge under the Digital Services Act. Additionally, the US Securities and Exchange Commission introduced a five-year exemption for trading tokenized stocks, aiming to bring blockchain-based shares closer to mainstream markets while ensuring they retain traditional rights like dividends and voting power. The technology sector is also seeing significant developments in autonomous transportation and grocery delivery partnerships. Lucid and Bolt are planning a major rollout of robo-taxis across Europe, with Bolt intending to deploy at least 25,000 autonomous Lucid vehicles using Nvidia's Hyperion platform by as early as 2028, though this remains a memorandum of understanding rather than a firm order. Meanwhile, Costco has expanded its delivery ecosystem by allowing DoorDash and Uber Eats to operate within its US warehouses, significantly increasing the number of states where these services are available and providing members with more app choices. This strategic shift gives delivery platforms a crucial foothold in the grocery market, ending Costco's exclusive long-term partnership with Instacart. Finally, OpenAI continues to face intense scrutiny from the mathematical community and legal challenges. Reports suggest that OpenAI employees expect the company to soon crack the Hodge conjecture, one of the Millennium Prize problems, which has sparked tension over concerns that AI labs are chasing headline-friendly proofs before the math community can verify them. Amidst this controversy, a federal judge denied OpenAI's request to review Apple's confidential agreement with Elon Musk's companies, ruling that the documents were not relevant to their ongoing antitrust dispute. As these legal battles continue, OpenAI is reportedly considering how to collaborate with mathematicians to avoid further backlash while navigating its complex relationship with both the tech industry and the academic world.

Read the full video transcript

[music] These are the daily tech headlines for Friday, September 18th, 2026. I'm Sarah Lane. The Wall Street Journal reports a threeperson security team at startup Hackron AI used Anthropics Claude to break into OpenAI systems through a bug in Open AI's discourse-based staff forum. Hactron AI was working through OpenAI's bug bounty program, gained access to employee chat GPT accounts and a GitHub software repository, then reported the issue and received $6,500. Open AAI says it fixed the flaws, but this is yet another example of AI tools making sophisticated hacking cheaper and faster. Newly unsealed court filings from news organizations suing Microsoft and OpenAI include some pretty blunt internal warnings. Microsoft Applied Science Director Brent Heck reportedly called AI scraping an astonishing theft and possibly the largest theft of labor in human history. Open AAI's Nick Turley warned publishers faced an existential threat from chatbot products trained on news. Microsoft says Brent Hex's comments were one employees view, not the company's legal position. Lucid and Bolt are planning one of Europe's biggest robo taxi pushes with Bolt looking to deploy at least 25,000 autonomous Lucid vehicles across major European cities. The vehicles would use Lucid's upcoming midsize platform and Nvidia Hyperion and Bolt would own and operate the fleet. Important caveat here. This is still a memorandum of understanding. It's not a firm vehicle order, but Lucid CEO says the first batch could arrive as early as 2028. A German court ruled Meta can be held liable for fake investment ads on Facebook and Instagram, ordering the company to remove the content, pay damages, and disclose information about the ads and revenue. The case involved a German financial portal and its founder whose logo and image were used in allegedly fraudulent investment posts. The court said Meta can't rely on the Digital Services Act's lack of knowledge defense because its algorithms and ad systems help control what users see. Meta obviously disagrees and is considering next steps. The US Securities and Exchange Commission or SEC rolled out a 5-year exemption for trading tokenized stocks, which could bring blockchainbased shares much closer to mainstream markets. Platforms will get relief from many exchange rules, but companies have to be notified before tokenized versions of their stocks are listed, and synthetic tokens that only mimic stock exposure won't be allowed. The SEC says tokenized shares have to carry the same rights as traditional securities. including dividends and voting rights. Open AAI launched Astra for law, a legal research configuration of GPT6 Astra for law firms and legal software companies. It's not a new model. It's a version wrapped with a legal search index covering US case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs. Initial access is limited to selected firms through a trusted access program in chat GPT and Codeex with an API version set to come later. Costco is giving Door Dash and Uber Eats a much bigger place in its delivery business. Door Dash can now deliver from Costco's US warehouses for members who link their accounts. And Uber Eats expanded Costco delivery from 17 states to 47. Costco has worked with Instacart for almost 10 years. So, this gives members more app choices and gives the delivery platforms that precious Costco foothold in grocery. Yes, everybody, we've got more OpenAI news. The information sources say Open AAI employees expect the company to soon crack the Hodgej conjecture. That's another one of the Millennium Prize problems. That's landing in the middle of OpenAI's fight with mathematicians over its Navier Stokes claim and concerns that AI labs are chasing headline friendly proofs before the math community can verify or even understand them. The report says OpenAI may wait to announce anything while it figures out how to work with mathematicians and avoid another backlash. And finally today, a federal judge denied OpenAI's request to see Apple's confidential agreement with Elon Musk's Excorp and Space X AI. Musk's companies recently resolved and dropped their antitrust claims against Apple, but are keeping claims against Open AI alive. And Open AAI argued the Apple agreement could matter to its defense. Judge Mark Pitman reviewed the materials privately and said they weren't relevant to the remaining summary judgement or trial issues. [music] For more analysis of the tech news of the day, subscribe to [music] daily technewshow.com. You can find show notes and links to all these headlines there as [music] well. I'm Sarah Lane. Thanks for listening. We'll talk to you tomorrow.