AI Valley
Posts
Hacking OpenAI with Claude

Hacking OpenAI with Claude

PLUS: 3 SpaceXAI employees are building a company in 3 days

Barsee
18 Sep

_{Sign up}_|_{Follow us on X}_|_Sponsor

Together with

howdy, it’s Barsee again.

happy friday, AI family, and welcome back to AI Valley.

here are the biggest things worth knowing today:

A 3-person team used Claude to hack OpenAI
3 SpaceXAI employees are building a company in 3 days
Plus trending AI tools, posts, and resources

Let’s dive into the Valley of AI…

ADOBE

New: Adobe Acrobat helps you unlock your own documents with new AI launch

Courtesy: Adobe

Acrobat's new AI helps you understand documents faster, turning dense reports into visual summaries, interactive reports, podcasts, and audio experiences you can take anywhere. When it's time to share your work, Stylize transforms plain documents into polished deliverables in just a few clicks. Every answer includes clickable citations, and Adobe does not train on your document data.

See it for yourself with Acrobat Studio today.

_{*This is sponsored}

THROUGH THE VALLEY

Courtesy: Hacktron AI team

A 3-person team used Claude to hack OpenAI - researchers at Hacktron AI used Claude Opus 4.8 and Opus 5 to break into OpenAI, eventually getting access to employees’ ChatGPT accounts and a private code repository. Opus 5 apparently produced the reliable exploit, and the whole thing took less than 72 hours. They didn’t download the code, reported everything to OpenAI, and got a $6,500 bug bounty.

The path started on community.openai.com. They found a year-old bug in the software used to process HEIC/HEIF images that gave them remote code execution on the forum. Then they discovered forum session tokens also worked on ChatGPT and Codex, including staff accounts connected to GitHub.

Claude did the hard part. Opus 4.8 found the bug but couldn’t build a reliable exploit. Opus 5, released July 24, got one working within hours. Hacktron says the whole thing took days of agent time, a few hours of human time and less than $3,000 in tokens. OpenAI fixed the SSO issue in ~14 hours.

Also this week, OpenAI published some weird things its own models did during training. An unreleased Astra model rewrote its notes to say it didn’t answer to corporations or governments, while GPT-5.6 Sol left instructions to invent missing data and hide errors.

So AI is getting very good at finding security holes. Keeping an eye on what the models themselves are doing is becoming part of the job too.

PRESENTED BY ATLASSIAN

Connect. Learn. Lead. Atlassian State of AI SDLC Digital Summit, September 22

Courtesy: Atlassian

Join Mike Cannon-Brookes (CEO & Co-Founder, Atlassian), Guillermo Rauch (CEO, Vercel), and leaders from Lovable, Dropbox, and more at Atlassian’s State of AI SDLC, a digital summit for engineering and product leaders on September 22nd. Together, they'll unpack how organizations need to adapt for the AI era. Reserve your spot.

_{*This is sponsored}

TRENDING TOOLS

Exa Snapshot - an index of 400 billion historical snapshots of webpages that lets you search as if it's the past
Astra for Law - OpenAI's frontier model GPT-6 is now available specifically for legal use cases
Muse for Mac - it works across apps, files, calendar, notes, and messages on your computer. You control what it can access. The team is shipping fast
Gemini 3.8 Live - Google's new voice models that think while they talk
Jev - a model intended to interact with other machines, rather than humans
Reception by ElevenLabs - an AI receptionist platform for small businesses, built on ElevenAgents
Arrow 2 - most advanced models for generating precise, editable vector graphics
Pion - agents for running fully autonomous companies
Mercury Books - Mercury's new bookkeeping AI for financial questions
Snap Specs Intelligence - AI assistant that learns your goals, priorities, relationships, and routines
ChatGPT is now in Word - you can add ChatGPT to Microsoft Word to turn rough notes into first drafts, untangle tricky paragraphs, proofread, and get suggested edits
Bend 2 - a new programming language that blocks AI mistakes via proof checking - the same technique big AI labs used to solve open math problems, like Navier-Stokes
Claude Cowork and chat are merging into one Claude - ask a quick question or hand over a report, and Claude takes it from there, even after you close your laptop
Salesforce in Claude - it brings your accounts, opportunities, and pipeline into Claude, with 37 pre-built sales skills

WHAT I'M CONSUMING

Hacking OpenAI
Three SpaceXAI employees are building a company in 3 days with Grok Bot
Chinese researchers detail 'the last AI built by humans'
The age of wonders and terrors
How to write with an LLM
Figure said Helix 2.5 completed whole-body household chores zero-shot across 30 unseen homes
Trump invites Tim Cook, Sam Altman, Jensen Huang, & other tech leaders to the White House state dinner for China’s Xi Jinping
Inside OpenAI’s agentic software factory.
OpenAI discloses six new incidents of ‘concerning' AI behavior
Manus said to close $500M funding at $4B valuation after Meta deal collapse
AI expected to cause net job losses in 34 of 37 countries surveyed by Pew Research