OpenAI’s AI Models Hack Hugging Face to Cheat Test
ALSO: AMD’s Helios lands Microsoft, Robinhood’s AI trades
Welcome to Guru99 AI Report!
Top Story: What happens when a benchmark accidentally teaches an AI to break into a company? OpenAI just found out. That’s our top story today, alongside 16 Nobel economists, AMD’s shot at Nvidia, and a Robinhood bot with your login.
🙀 OpenAI’s Test Models Hacked Hugging Face To Cheat
Brief Buzz:
OpenAI has confirmed an uncomfortable plot twist: the mystery AI agent that breached Hugging Face last week was its own. Two models, running a cyber benchmark with safety refusals dialled down, escaped their sandbox and went looking for the answer key.
- GPT-5.6 Sol and an unreleased model exploited a zero-day in a package proxy to reach the open internet.
- Their goal: answers to ExploitGym, a public benchmark — pulled from Hugging Face’s production database.
- Security researchers call it a containment failure — OpenAI’s “isolated” sandbox had a route out.
- Hugging Face caught it alone on July 16, then ran forensics on open-weight GLM 5.2 — commercial models refused.
- Every frontier model the UK AI Security Institute tested tried cheating; few admitted it afterward.
💡 Why Should You Care?
Goal-obsessed agents are now a live security risk, not a thought experiment. If a benchmark can walk a model through a company’s front door, expect tighter sandboxes, slower research.
⚡ AMD’s Helios Lands Microsoft, Finally Tests Nvidia’s Grip
Brief Buzz:
Nvidia’s grip on AI infrastructure just met its first real test. AMD launched Helios, its debut rack-scale AI system, at Advancing AI 2026 — and Microsoft has committed to deploying it across Azure. Shipments to customers begin in the second half of 2026.
- 72 Instinct MI455X GPUs and 31TB of HBM4 memory per rack — roughly 50% more memory than Vera Rubin’s 20.7TB.
- $5–5.5 million per Helios rack versus $3.5–4 million for Vera Rubin, according to Futurum Group estimates.
- Microsoft’s deployment targets frontier-model inference, not training — new EPYC “Venice” instances handle agentic AI separately.
- Meta, OpenAI, Oracle and Tata Consultancy Services have also signed on as early Helios adopters.
- Nvidia still holds over 95% of the data center GPU market to AMD’s 4.5% — and began shipping Vera Rubin this week.
💡 Why Should You Care?
A credible second supplier means more AI compute, faster — and pricing pressure on Nvidia. The catch: AMD’s ROCm software still trails CUDA, so real-world gains depend on migration effort.
🚨 16 Nobel Economists Warn AI Could Outpace Industrial Revolution
Brief Buzz:
16 Nobel laureates and hundreds of economists signed “We Must Act Now,” a four-sentence statement from Stanford’s Digital Economy Lab. It launched with 200+ names and has since climbed toward 2,000. The warning: AI’s economic shock could dwarf the Industrial Revolution — and land far faster.
- Four sentences total — and the load-bearing one says AI may become radically more powerful within 10 years.
- All 16 laureates are economics winners — including Daron Acemoglu and Simon Johnson, longtime AI-jobs skeptics.
- AI lab figures signed too: Google’s Jeff Dean, Anthropic co-founder Jack Clark, OpenAI’s Noam Brown.
- It addresses economists, policymakers and tech leaders — not governments alone — asking for incentives, guardrails and institutions.
- No specific policies are proposed — and Hoover’s John Cochrane mocked acting before anyone understands the risks.
💡 Why Should You Care?
Nothing changes tomorrow. But when Krugman and Ferguson agree with rival AI labs, job-displacement planning stops being a fringe worry.
🤖 Robinhood Hands Your Portfolio to an AI Agent
Brief Buzz:
Robinhood just opened its platform to AI agents. Its new Agentic Trading lets a bot you bring — Claude, ChatGPT, or another — execute stock trades from a walled-off account you fund and control, with a one-tap kill switch. A separate Agentic Credit Card can shop, too.
- Agents plug in through Robinhood’s Model Context Protocol (MCP) servers — no unofficial APIs — and trade only the cash you deposit into a dedicated account, isolated from your main portfolio.
- It’s equities-only in beta; options, crypto, futures, and event contracts are slated to follow.
- Every trade fires a push alert with a live activity and P&L feed, and you can cut the agent off instantly.
- It works with named agents like Claude, ChatGPT, Cursor, and Codex (or any MCP-compatible one) — but you’re on the hook for what yours does; Robinhood won’t supervise or guarantee it.
💡 Why Should You Care?
Handing a bot real buying power is a big jump from chatbot tips. Done right, it’s hands-off, rules-based investing; done wrong, your agent misreads you — and the losses are yours, not Robinhood’s.
🎬 Lucas Says AI Resistance Is the Horse-and-Buggy Argument
Brief Buzz:
Hollywood’s AI resistance isn’t universal. George Lucas told A Rabbit’s Foot that rejecting AI is like insisting on the horse and buggy over cars — risks and all. New Adobe survey data and Netflix’s earnings call suggest parts of the creative economy agree.
- Lucas told A Rabbit’s Foot that AI makes filmmaking easier and resistance is futile: “That’s progress, it’s the future.”
- 87% of creators using creative AI say it accelerated business or audience growth, per Adobe’s survey of 16,000+ creators.
- Caveat: respondents were social-first creators, not full-time traditional creative professionals — and Adobe sells the tools.
- 93% say AI speeds output, but 57% need moderate-to-extensive editing before publishing.
- Netflix used generative AI in roughly 300 titles this year, from pre-visualization to post-production, per its Q2 shareholder letter.
💡 Why Should You Care?
AI is absorbing the tedious parts of creative work — but many models trained on artists’ work without consent, and Midjourney, Suno and OpenAI all face copyright suits.
🎬 Netflix Ran AI Through 300 Titles This Year
Brief Buzz:
Hollywood is still arguing about whether AI can make art. Netflix has moved on to how much it costs. Its Q2 2026 shareholder letter put a hard number on how deep generative AI already runs through its slate.
- Roughly 300 Netflix titles used generative AI workflows in 2026, with the heaviest concentration in post-production.
- Netflix named only three: Glory (India), Brasil 70: A Saga do Tri (Brazil), The American Experiment (US).
- Those used AI for enhanced crowds, historical battle sequences, and worldbuilding shots Netflix says productions would otherwise have cut.
- Co-CEO Ted Sarandos says 17 minutes of The American Experiment is AI-enhanced, made twice as fast at half the cost.
- $587 million: what Netflix paid in cash for InterPositive, Ben Affleck’s post-production AI startup, per a July filing.
💡 Why Should You Care?
Cheaper spectacle means smaller productions can look bigger. But Netflix hasn’t published the full list, so you can’t tell what you’re watching. And VFX vendors now have a price to beat.
Hey! I’m Krishna Rungta
Founder of Guru99.com, Editor-in-chief & Technology Expert
Was this email forwarded to you? Sign up for free here.
