{{current_date_full_with_abbr_day}}
3-minute read
3
AI UPDATES
2
TOOLS
1
READ
👋 Good morning.
The old Claude couldn't hack OpenAI. The new one did it in hours. A three-person team called Hacktron used Claude Opus 5 to turn a booby-trapped iPhone photo into access to OpenAI employee accounts, per TechCrunch. Then they told OpenAI, which fixed the hole and paid them a $6,500 reward for finding it. Anthropic's model, OpenAI's cheque.
IN 3 MINUTES, GET 1% BETTER THAN EVERYONE YOU KNOW:
54% would let an agent apply for credit
OpenAI built a legal index of 230 million pages
Runway fixes your frame rate for pennies
Pay AI agents only when the work passes
What 80,508 Claude users want, and fear
3 AI updates
1. 💳 Trust, but bank
Experian had Forrester survey 6,247 credit-active consumers across 13 markets in Europe, Africa and Asia Pacific, per its September 16 release. 54% are comfortable letting an AI agent apply for credit for them, and 75% would feel better if the agent were tied to a bank they already trust. 82% already trust AI to compare loans. But only 23% would let it act under pre-agreed rules, and 5% want full autonomy. The job people want done is the comparison, not the signature.
Do this: If you sell anything with a checkout, let the agent compare and prepare, and keep a human approval step before money moves.
Half want the agent. Three-quarters want the bank holding its leash.
Would you let an AI agent apply for credit for you?
2. 💼 Law without tab soup
OpenAI launched Astra for Law on Thursday. It is not a new model. It is GPT-6 Astra wired to a new US legal search index of more than 230 million pages of case law, statutes and regulations, including Free Law Project's collection covering over 99.9% of published precedential cases. On 200 questions from Vals AI's legal research benchmark, it passed 54% against 38.7% for Astra with plain web search. Selected US firms only for now, through ChatGPT and Codex.
Do this: If your lawyer bills you for research hours, ask which tool they use now. 54% is a pass rate, not a guarantee, so every citation still gets checked.
230 million URLs later, the tabs finally clock out.
Runway shipped Enhance Frame Rate on Thursday. It converts any video, made on Runway or not, to 25, 30, 48, 60 or 120 fps, including broadcast's 59.94, without changing the resolution. It handles up to 4K and five minutes, at one credit per two seconds no matter the target. Runway says it is up to seven times faster and three times cheaper than other interpolation models, which is Runway's own benchmark.
Do this: Run your choppiest 24 fps clip up to 60 fps once. A one-minute clip costs 30 credits.
Your 24fps draft just learned to speak broadcast.
2 Tools worth knowing
1. 💰 Bounty · pay per verified task · web
Post the outcome you want, set how it gets checked, and fund a reward. AI agents compete for it, an automated checker grades the work, and the money leaves escrow only if it passes. 200-plus agents, 2,000-plus tasks done so far, per the site. The catch: you pay into escrow before work starts, and a vague brief gets you a vague result.
2. 🧠 TypeSafe Jev · early access · API
TypeSafe, a startup from ex-OpenAI researcher Diogo Almeida, just launched Jev, a model that only makes decisions. You give it a list of possible answers, and it picks one and tells you how confident it is. Responses land in 70 to 500 milliseconds, while chat models take seconds. The flex: Jev plays Doom by reading the game as text, making about 10 moves a second for roughly $7 an hour. The catch: every speed number comes from TypeSafe itself.
1 Read
Anthropic interviewed 80,508 Claude users in 159 countries and 70 languages about what they want from AI and what scares them. Top wants: getting better at their job (18.8%), personal change (13.7%), running their life (13.5%). Top fears: that it's unreliable (26.7%), jobs (22.3%), losing control over their own decisions (21.9%). 67% came out net positive, and hope and worry usually sat in the same person.
The catch: these are people who already use Claude, surveyed in March, not the general public. Read it as what your most engaged customers worry about.
81,000 opinions. One shared nervous system.
Don't forget to rate today's post
That's your 1% for today. Reply with the one thing you'd never let an agent sign for you.
Kushank, AI teammates, and the human team
PracticalyAI
