Researchers used Claude to hack OpenAI

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

🕵️ Researchers used Claude to hack OpenAI

🛰️ SpaceX want to buy dead startups' data

🤖 AI titans invited to Trump-Xi dinner

🦹 Microsoft exec called AI scraping 'theft'

🧠 Claude leads 26% of Anthropic's AI R&D

Plus: 🎁 14 other news you might like, 🧰 6 tools, and 📚 5 papers.

🕵️ Researchers used Claude to hack OpenAI LINK

Independent researchers at startup Hacktron AI used Anthropic's Claude to break into OpenAI, chaining two serious flaws to reach several employee ChatGPT accounts and earning a $6,500 payout through OpenAI's bug-bounty program.

The way in started with an image upload on OpenAI's community forum, which runs Discourse, hitting a memory bug in the libheif library that let attackers sneak in their own instructions and take over the server.

The Claude version Opus 4.8 failed to build a working exploit across several tries, but within hours of Anthropic releasing Opus 5, the same model cracked the problem, unsettling researchers about how fast AI is speeding up hacking.

🛰️ SpaceX want to buy dead startups' data LINK

SpaceX is quietly in talks to buy customer and operational data from failed or struggling startups, giving it cheap AI training material, though sources say the discussions are informal and may never turn into actual deals.

The data could help train Grok models from xAI, the SpaceX-owned company whose systems trail Anthropic, OpenAI, and Meta; the approach mirrors Google buying Spirit Airlines business records for $10 million at a bankruptcy auction last month.

xAI runs Grok 4.6 now, plans Grok 4.7 this month, and expects Grok 5 later this year, which Musk said will fold in every piece of data SpaceX has produced across its 25-year history.

🤖 AI titans invited to Trump-Xi dinner LINK

Several top AI and tech leaders have been invited to a dinner tied to the September 24 summit between President Trump and China's Xi Jinping, including Qualcomm's CEO, who a spokesperson confirmed will attend.

The gathering lands as parts of the AI industry push both leaders to set limits on fast-moving model development, with OpenAI's Sam Altman backing Anthropic chief Dario Amodei's call for an industrywide slowdown and shared safety standards.

Nvidia's Jensen Huang, who has called fears of AI ending humanity "hysterical," has pressed the White House and Congress to let Nvidia sell its top chips to China, raising the question of whether chip sales will surface in the talks.

🦹 Microsoft exec called AI scraping 'theft' LINK

Newly unsealed filings in The New York Times' copyright case against OpenAI and Microsoft show a top Microsoft executive privately called the companies' AI training practices "theft" and warned publishers faced serious harm.

Microsoft's own director Brent Hecht described the copying as "the largest theft of labor in human history," while internal data showed the Copilot answer engine cut click-throughs to the Times' site by as much as 93%.

The documents claim OpenAI's mid-training datasets held over 91,000 copies of works from the news outlets, that staff planned to slip past paywalls undetected, and that they stripped copyright notices before training.

🧠 Claude leads 26% of Anthropic's AI R&D LINK

Anthropic says its Claude chatbot now runs 26% of the company's AI research and development, meaning it handles those jobs almost start to finish from a broad prompt while a person keeps watch over the work.

The company added that AI does at least large parts of the work under close human direction on over 90% of its research, and stressed Claude "is not operating fully autonomously" for any measured slice of R&D.

Anthropic shared the numbers in a blog post proposing three ways to measure AI progress, tracking Claude's "automation level" since August 2025 using a rating scale from Epoch AI that rival developers could copy with outside checks.

Other news you might like

• OpenAI launches Astra for Law, and its own legal research indexLINK

• Anthropic launches Claude Code Projects, an ‘always-on’ conversation that remembers and delegates your long-running dev workLINK

• CFTC Opens Door for Crypto Apps to Offer Regulated Derivatives AccessLINK

• Paris prosecutors open criminal probe into smart glasses used to film womenLINK

• UN turns to Google to make its global data ready for AI agentsLINK

• The FAA’s plan to fix air traffic? $875 million worth of AILINK

• Chinese AI companies are earning a fraction of what OpenAI and Anthropic make, study findsLINK

• Treasury Sanctions Crypto Exchange Behind Iran's Bitcoin Tolls on Hormuz ShipsLINK

• Australia considers banning the use of smart glasses in government buildingsLINK

• Waymo will launch robotaxis in Singapore in 2028, its first Southeast Asia marketLINK

• Palantir CEO Alex Karp calls for nationalizing frontier AI companies over ‘unlimited risks’LINK

• Anthropic has set up a bio research lab for physical experimentsLINK

• Jensen Huang says Nvidia will sell twice as many chips next yearLINK

🧰 Trending tools

cubicles.lol: a browser-based multiplayer metaverse where you claim and customize a cubicle to showcase projects, meet other builders, and play mini-games.LINK

Shall We Talk: converts speech to clean text on iPhone and Mac, dictating into any app and turning meetings into speaker-labeled transcripts and summaries.LINK

Second Eyes: sorts and shortlists your photos, explaining why each was picked and what needs checking, so you make the final publishing call.LINK

OpenAlgo Charts: open-source, data-agnostic charting library for brokers and trading platforms, letting you plug in your own market data without building charts from scratch.LINK

Banana Keyboard: an Android keyboard that inserts curated AI prompts with one tap into any chat app, keeping favorites and recents on your device.LINK

Snooze Files: temporarily hides files from Finder and returns them to their original folder at a set time, clearing desktop clutter without losing track of anythingLINK

📚 Trending papers & reports

Music-describing AI often invents confident details a song doesn't contain, and this study tests nine such systems, finds vocals are universally misjudged, names Audio-Flamingo-3 the most reliable, and shows quick fixes only partly help.LINK

Underwater robot self-repair gets a test bed showing that a top-tier language model correctly diagnoses a weight-shift fault ~85 to 90% of the time, versus ~60 to 78% for cheaper on-device models.LINK

Eye-care decision support pulls answers directly from medical guideline pages, images, tables, and all, boosting accuracy on the hardest clinical questions by ~24 percent over a top general model while showing doctors the exact source page.LINK

Robot task planning lets a robot draw several possible movement paths, have a vision model pick the safest and most efficient one, then execute it, hitting ~71% success versus ~50% for prior methods across eight real tasks.LINK

Heart-scan reading software matches specific squiggles in an ECG to individual diagnoses instead of judging the whole trace at once, and fills in missing report details using screened AI text, beating prior methods across six datasets.LINK

See you tomorrow for a new dose of ☕️ Techpresso!