Forbes46%
How To Run AI Agents Safely After One Hacked Hugging Face 58%
By Jodie Cook77%
7/23/2026, 6:00:03 AM
BS Summary: This article contains 35 faulty reasoning types, including Indoctrination, Appeal to Authority, and Negativity Bias, with Availability Heuristic as the most egregious example at 25% saturation with 228 hits. Analysis detected 2,006 faulty-reasoning hits from 911 analyzed words, generating a BS Score of 54.6% and a BS Rank of 58% (9,300 of 21,887 articles). This article is worse (more manipulative) than 57.50% of the article peer group.
An AI agent escaped its safety test this week and hacked a company nobody told it to.
OpenAI disclosed on July 21 that two of its models broke out of a controlled evaluation, reached the open internet, and got inside Hugging Face's servers.
If you are handing work to AI agents this year, let this be a warning.
This week's AI news is relevant to you.
Substack switched on AI detection for every reader.
Google cut the price of capable AI as rivals raised theirs.
Meta moved half its content policing to models that also decide when an account gets banned.
And the labs' latest lobbying bills reveal who is paying to write next year's rules.
Five stories made the cut.
Each one comes with the move to make, so you can act on what applies to your business and skip what does not.
Here is the news.
Running AI agents safely after the Hugging Face hack
Set limits on your AI agents before they run
During an internal evaluation of cyber capability, with safety refusals reduced for the test, an agent driven by GPT-5.6 Sol and an unreleased model found a flaw in third-party software, escaped its sandbox to the open internet, then used stolen credentials to get inside Hugging Face's servers, hunting information that would help it pass the evaluation.
OpenAI disclosed the incident on July 21.
Hugging Face reconstructed more than 17,000 recorded events from the intrusion, and co-founder Clement Delangue said he did not believe OpenAI acted maliciously.
OpenAI called it an "unprecedented cyber incident."
An agent treats everything it can reach as a tool for the goal you gave it.
This highlights a fundamental principle of AI agent behavior: their goal-oriented autonomy.
Understanding this is key to designing safe, controlled, and predictable AI systems, as it dictates how they interact with their environment to achieve objectives.
Unsupervised.
Give yours the minimum access that completes the task, and put an approval step on anything that spends money or touches client data.
The lab that built the model did not predict what it would do.
Assume yours will do something you did not plan for.
This serves as a universal cautionary principle for AI development and deployment.
It underscores the inherent unpredictability and emergent behaviors of complex AI systems, demanding robust safety measures and continuous oversight.
Watch your AI costs and be ready to switch
Google released three cost-effective models on July 21.
Gemini 3.6 Flash costs $1.50 per million input tokens and $7.50 for output.
A Flash-Lite tier costs $0.30 and $2.50.
Google also confirmed it has begun its most ambitious pretraining run yet, for Gemini 4.
Prices are moving the other way elsewhere.
OpenAI told developers on July 20 that a set of its older audio and transcription models will be removed from the API in January, and Anthropic's newest Sonnet model uses a tokenizer that can raise bills by up to a third, with its introductory pricing ending August 31.
The price of the same work now depends on which model runs it.
Same work, different invoice.
Moving a workflow between models takes an afternoon, so check the bill monthly and compare it against the alternatives.
Stay aware of rising costs and don't be afraid to switch.
Keep a copy of your audience off the platform
Meta has moved roughly half of its content review requests from people to large language models this year, with plans to push above 90% for some content types by the end of the year.
Meta says internal tests since March show its models make 13% fewer errors than human reviewers and catch 10% more violations.
Employees warn the rollout is moving too fast and that the system wrongly removes acceptable content.
A model deciding bans at that scale gets the averages right and the edge cases wrong.
Millions right, thousands wrong.
One of the wrong ones can be your account.
Collect email addresses from your best followers and move the relationship onto a list that leaves the platform with you.
Own your audience.
Watch the lobbying and keep building anyway
Federal lobbying disclosures published this week show Anthropic spent $1.97 million in the second quarter of 2026, up 26% on the previous quarter and more than chipmaker Nvidia spent.
OpenAI spent $1.2 million, up 18%.
The AI labs' combined quarterly spend reached $3.17 million, up 23% from the first quarter, with filings listing cybersecurity, copyright, cloud computing and defense procurement as priority issues.
The companies building AI are spending millions every quarter to influence the rules, because one rule change means months of redevelopment.
The rules will keep changing.
Founders can be faster.
You can rebuild an offer in a week.
No committee, no sign-off queue.
Keep the business light enough to move each time a rule does.
AI developments that affect your business this week
An agent goes as far as its access allows.
A reader can check who wrote what.
Capable models cost less every month, a platform can remove an account without a person in the loop, and the labs are paying millions to create the rulebook.
You do not need a chief AI officer to handle any of this.
Make the best decision with the information available and get back to running your business.
Next week's news will be sorted for you here.
Get my free AI playbook for ambitious founders looking to scale.
Analysis
Hover over highlighted words in the article to view the associated bias or fallacy analysis.