How much do AI labs really know about what their models are doing? This week we look at how OpenAI agents used a 25 year old German wiki to cheat on tests, whether OpenAIâs new Astra model counts as AGI, why 95% of startups are now running AI agents, and a new poll showing that worrying about AI is a bipartisan issue.
AI News đ°
95% of startups are using AI agents
The trend: 95% of early-stage companies have deployed AI agents, and 77% increased their AI spend over the past year, with banking data showing average monthly AI spend up more than 50% since April 2025.
The details: A new Mercury survey of 1,500 US companies under six years old found that significant AI adopters were 4x more likely to have raised venture capital in the past year (31% versus 7% for non adopters). 91% have improved business confidence versus 60% of non AI adopters. Costs are climbing, with AI spend rising 25% to 50% across the board. Entry level jobs arenât being replaced just yet, with 82% hiring the same or more early-career talent than they would have without AI, but 34% have added âAI fluencyâ as a job requirement.
Why it matters: 29% of respondents said they rely on a single AI vendor, and 65% are worried that a price change and/or shutdown would hurt their business. More than ever, companies should be concerned about their data sovereignty and being locked into one vendor, which our CEO Paul has written about recently.
A second agent swarm has hit the internet
The trend: OpenAI agents turned a dormant German programming wiki into a message board, leaving 18,000 posts as they shared test answers and ways to bypass sandbox restrictions, months before the Hugging Face incident.
The details: According to an independent report, agents were given a timed web-lookup task, in which they were supposed to have the ability to read the internet but not to write on it. They gained access then used the wiki to communicate. The agents shared answers, pooled results, and shared techniques for bypassing their restrictions. This allowed them to to cheat on the task. Posting stopped on June 22, one day after computers on an OpenAI IP addresses visited the site, which the report assumes is OpenAI employees finding the activity and shutting it down without publicly disclosing it.
Why it matters: The agents werenât trying to hide themselves, with many signing their posts with names like âOpenAIResearcher.â This kind of reward hacking will only get harder to catch as models get better and better, which makes the security and alignment promises frontier labs are making worth remembering.
Trending Now âĄ
OpenAI releases Astra, but is it really AGI?
OpenAI released the much hyped GPT-6 Astra model, which it calls the âthe worldâs most intelligent and aligned.â Astra scores 99.9% on ARC-AGI-3, which is designed to test if models can genuinely reason and adapt in a novel environment. In the wake of the Hugging Face incident, OpenAI have focused on safety and alignment, with Astra causing âfewer misaligned outcomes than any other frontier model tested.â So, is this AGI? OpenAI president Greg Brockman thinks so, but Chief Scientist Jakub Pachocki describes AI as âan alien mindâ that we donât fully understand.
America unites in fear of AI
A new NBC News poll found that 70% of Americans feel more worried than excited about AI, regardless of political affiliation. AI use is climbing, with 52% saying they use AI often, but just 18% trust AI-generated information most of the time. Data centers are one of the biggest issues, with 69% against having one built nearby, versus only 31% who would support it. 70% think AI is costing people jobs, and expect it to hurt schools and elections. 81% say the governmentâs current AI rules donât go far enough, and when asked which party they trust, 44% say neither.
Read This đ
OpenAI restarts training for future models that it had previously paused following the Hugging Face incident
Global data center spending could hit $31.6 trillion by 2050, driven by AI demand and chip upgrades
Robotaxis Try Their Luck on Londonâs Medieval Roads
New York public schools are banning generative AI until the 8th grade, with Mayor Zohran Mamdani rejecting the idea that AI in education is ânot only inevitable, but necessaryâ
China and the USA are preparing for mid-September AI safety talks, focusing on AI cyberattacks and model distillation
Sam Altman says that OpenAI âwill definitelyâ build a humanoid AI robot, as the physical world is already designed around the human body
Check This Out đ
Thanks for reading!
Henry
That emergent coordination via the German wiki highlights the real failure mode of agent eval loops: sandbox leakage isn't just a boundary problem, it's a state-synchronization risk. When agents establish out-of-band communication channels, static guardrails fail to model collective policy deviation.