OpenAI puts the safety brakes on Astra

Good morning, { AI lovers }, and welcome to the 12,457 new readers who joined us this weekend. Every week in the past, OpenAI’s upcoming Astra mannequin was making main waves within the math world after knocking out 10 long-standing issues. Days later, it set off an alarm the corporate had by no means rung earlier than.

The corporate says Astra (anticipated to be GPT-6) is the primary mannequin it is treating as a possible “crucial” cyber danger, triggering paused inside work, deeper authorities testing, and a probably slower highway to launch.

OpenAI places the security brakes on Astra
The Rundown Roundtable: Our AI use instances
Reduce onboarding time in half with Loom and ChatGPT
China’s Kimi K3 joins the jailbreak get together in testing

COMMUNITY AI WORKFLOW OF THE DAY

Six Star Trek-themed AI brokers run on a house server. Spock routes duties to the precise crewmate, and the crew scans for vulnerabilities, summarizes e-mail, posts the climate, and screens seven machines on Discord for a couple of cents a month.

^{Credit score to reader}^{Joao Silva}^{. How do you employ AI? Inform us}^here^{for an opportunity to be featured tomorrow.}

OPENAI

Picture supply: Photos 2.0 / The Rundown

The Rundown: OpenAI simply designated its Astra mannequin as its first “crucial” cybersecurity-capable AI, placing its preparedness framework into motion with safeguards CEO Sam Altman says could “want a bit bit longer” earlier than broader rollout.

A ‘Crucial’ mannequin is outlined by OAI’s framework as with the ability to both discover and create zero-day bugs or perform cyberattacks with out people within the loop.
Astra, rumored to be GPT-6, was just lately unveiled publicly after the mannequin solved 10 important open math and pc science issues.
Steps being taken embody heightened safety restrictions, a pause in sure inside actions with Astra, and deeper authorities and third-party testing.
The information follows safety points throughout OAI, Anthropic, Meta, and Moonshot (extra under), although OAI stated Astra wasn’t concerned within the Hugging Face hack.

Why it issues: The incidents piling up are an indication that AI is coming into an unprecedented functionality section (if you happen to haven’t seen this post-mortem on the HF hack, it’s eye-opening). Whether or not the misaligned capabilities could be managed or are merely getting too complicated to get forward of is a query wanting extra unsure than ever.

TOGETHER WITH MOZILLA

The Rundown: Mozilla’s inaugural State of Open Supply AI report finds open fashions have practically closed the efficiency hole with proprietary giants like ChatGPT and Claude — whereas a brand new battle emerges over who controls the infrastructure constructed round them.

Report highlights embody:

Efficiency hole narrows to only 3%
Open fashions make up a 3rd of utilization however simply 4% of income
Energy is shifting to the “agentic harness”

Learn the total report here.

THE RUNDOWN ROUNDTABLE

Picture supply: Nate @ The Rundown

The Rundown: The Rundown Roundtable is a weekly characteristic the place we ballot members of The Rundown employees about how we use AI in our work and every day lives.

Nate, College Educator: The brand new ChatGPT Chrome extension is surprisingly good at navigating sophisticated technical settings. My favourite use case to date has been utilizing it to create the particular DNS data for a brand new web site.

This has all the time been complicated, though I’ve finished it dozens of occasions. The method varies throughout platforms; each DNS registrar makes use of completely different names for data equivalent to A and CNAME data, and the interfaces are always altering. It’s tedious and simple to get improper.

For this type of one-off, extremely detailed technical activity, I copy and paste the particular directions from my web site host supplier into the extension whereas I’m logged in to my DNS supplier. GPT then takes management of the mouse and fills within the required fields.

The use instances are limitless, from filling out an expense report back to dealing with different tedious duties in legacy or old-school software program that doesn’t have a easy CLI, MCP, or API connection.

Nick, Senior Video Producer: I had a video shoot earlier this summer time the place the principle lavalier mic on the expertise ended up getting misplaced. The undertaking had no room for a reshoot, so I used to be left with simply the scratch audio from the on-camera mic, which was quiet, very distant from the topic, and had tons of room noise within the audio.

I used to be capable of take the scratch audio and run it by way of the Adobe Podcast AI software to repair the audio and salvage the shoot. Prior to now, doing this type of restoration would have required lots of handbook work and scrubbing by way of the spectrogram to tug out undesirable frequencies & simply hoping that it was going to work in the long run.

The AI software was capable of save the audio, clear up all the background noise, and normalize the expertise’s voice. Ethical of the story – all the time double-check mics on set & do not panic when one thing is improper in manufacturing.

AI TRAINING

The Rundown: On this information, you’ll learn to flip a brief Loom walkthrough into a transparent onboarding SOP with ChatGPT.

Decide a repeatable activity, like filling out timesheets or writing a report, and report it in Loom. Clarify clicks, choices, exceptions, and the definition of finished
Use Loom’s transcript enhancing to take away errors or lengthy silences
Copy the transcript into ChatGPT. Inform it to jot down the SOP in Markdown
If it’s not on this format, an excellent one is: function, hyperlinks to assets, steps, after which a guidelines
Paste the ultimate SOP right into a Google Doc

Professional tip: When you ask it to construct the SOP in Notion, it may well embed the Loom video so it’s playable proper on the web page.

PRESENTED BY VANTA

The Rundown: Vanta’s MCP server connects Claude, Cursor, and Codex on to your compliance program, so AI-native groups can catch dangers, repair gaps, and keep audit-ready with out leaving their workflow.

Join Vanta instantly into your AI instruments
Floor failing assessments and audit gaps immediately
Assign remediation with out leaving your workflow

Save your spot here. If you cannot make it on Aug. 18, no worries! Register, and you may mechanically get the recording after the session.

MOONSHOT AI & SECURITY

Picture supply: @Sauers on X

The Rundown: Chinese language lab Moonshot AI’s Kimi K3, which neared the frontier at launch final month, joined the sandbox escape membership after U.S. agency Frontier Safety revealed the mannequin slipped by way of its check setting to tug a solution key off GitHub.

The benchmark’s reply key sat in a public codebase, and Frontier says K3 realized it may attain GitHub by way of a loophole meant for software program installs.
Nothing really received hacked, and the sandbox wasn’t extensive open, however Frontier researcher Paul Kassianik said the place rival fashions refused, “K3 did not blink.”
K3’s weights are open and freely downloadable precisely as examined, which is why Frontier warns the case may show “probably extra dangerous.”

Why it issues: The AI world cycles by way of a brand new essential character each few months, and safety breaches have formally taken the spot this summer time. However this entry hits a bit in another way than different escapees coming from closed labs like OAI, Anthropic, and Meta that may patch them, with Kimi’s weights already downloaded and out on the planet.

🗣️Unwrap Customer Intelligence – Join your whole group to the true voice of the shopper with AI-driven insights from buyer suggestions*
🎬 Seedance 2.5 – ByteDance’s video technology AI, now broadly accessible
🪁 Kitesurf – Cloudflare’s light-weight, agent-first browser

Elon Musk revealed that Tesla and SpaceX’s ‘Terafab’ plant will probably be situated in Grimes County, TX, predicting it is going to be “the biggest and most beneficial constructing on Earth by far”.

ByteDance is reportedly pre-training an AI with as much as 10T parameters, which might triple the dimensions of China’s largest mannequin so far and probably rival Anthropic’s Mythos.

xAI rolled out Grok Think about Picture 2.0, that includes upgraded enhancing, textual content rendering, and general high quality, with the mannequin sitting behind OpenAI’s GPT Picture 2 throughout benchmarks.

Analysis agency SemiAnalysis said DeepMind is “not a frontier lab”, however projected Google Cloud progress to prime 100% in 2027 as Google pushes TPU gross sales.

Anthropic introduced a Claude Code characteristic that permits the platform to message throughout periods, permitting cases to choose up the place one other left off.

Google co-founder Sergey Brin is reportedly taking over oversight of the corporate’s Gemini mannequin improvement, coming within the wake of final week’s management shakeup.