In partnership with
Reading time: 6 minutes
🗞️ In this edition
Claude Broke Into OpenAI
Alibaba Made Video Cheap To Watch
Sponsored: Attio
Palantir Hired Its Way Into Westminster
⚡ YOUR MOVE For people building the life they want, now.
In other AI news –
Adecco Put An Agent In Front Of 27,000 Staff
Google Taught Its Agents To Stop Repeating Themselves
Insurers Started Writing AI Out Of The Policy
4 must-try AI tools
Three people reached OpenAI's internal code repository, and the thing that got them in was Claude. Alibaba cut the price of watching video by 89%. And Palantir hired a former deputy leader of the Labour party.
Keep reading.
Source: TechCrunch
WHAT HAPPENED
Researchers at Hacktron AI, a three person security startup, chained a heap buffer overflow in the libheif image decoder to a flaw in OpenAI's single sign on and reached the company's internal monorepo, where they opened pull request #1186742 as proof. The way in was community.openai.com, OpenAI's Discourse forum, which converts uploaded iPhone photos using a library that had never received a Debian security backport. Work started on July 23 with Claude Opus 4.8, which the team says struggled across several sessions to produce a working exploit with ASLR enabled. Anthropic released Claude Opus 5 on the evening of July 24, and the researchers had a working ARM64 exploit within about three hours and remote code execution on Discourse Cloud by 10am the next morning. OpenAI confirmed a fix at 22:49 UTC on July 25 and paid a $6,500 bounty on September 1, noting that testing against the Discourse hosted forum was excluded from its programme and the award covered the OpenAI side finding.
WHY IT MATTERS
The number worth holding onto is not the bounty. It is the gap between two model releases. Same three people, same target, same bug, and the thing that decided whether the attack worked was which version shipped that week. Every security budget written this year assumed offensive capability improves at the speed an attacker learns. It now improves at the speed a vendor ships, and the vendor ships to everybody at once, for roughly $200 a month. Gray Swan chief executive Matt Fredrikson put it plainly, that for $200 a month anyone can use these tools and hack into a company like OpenAI.
OUR TAKE
The obvious read is that this is a story about OpenAI's security, and it is not. OpenAI patched inside a day and the forum was out of scope anyway. What actually happened is that a frontier model release quietly became a patch deadline for every company on earth, and no disclosure process in existence runs on that clock.
Watch what ships alongside the next frontier launch. If no major lab publishes an offensive capability delta on release day before the end of this year, the industry has decided this is somebody else's problem, and the next writeup like this one will not arrive with a bounty attached.
From Our Sponsor:
Some teams never seem to stop moving. They're on Attio, the agentic CRM.
Every customer signal is captured in one shared context layer, always current and compounding. Agents and workflows build pipeline, chase every buying signal, and move deals forward, an always-on revenue engine running alongside your team.
With Attio, you’ll get:
Leads automatically prioritised and routed to the right rep
Expansion and risk signals caught the moment they land
Follow-ups written in your voice, already there when you arrive
Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?
Source: MarkTechPost.com
It is Qwen's first omni modal model built around agentic work, taking text, images, audio and video in and returning text, with a one million token context window. Qwen reports a 25% average gain across 29 evaluations against Qwen3.5-Omni-Plus, and says video input now costs about 89% less than that model, helped by an agentic perception mode that cuts token use by roughly half. Pricing is $0.15 per million input tokens and $0.47 per million out, with cache hits at $0.016. There are no open weights. It is API only at launch.
When you don't click through on a story, what's usually the reason?
Source: The Times
Watson was Labour's deputy leader, has advised Palantir since 2024, took a leave of absence from the House of Lords in March and surrendered his parliamentary pass to avoid a conflict of interest. He arrives with five months on the clock. Palantir's seven year Federated Data Platform contract with the NHS is worth £330 million and carries a break clause that opens in February 2027, at which point the government can walk if somebody else can do the job better or cheaper. Executive Louis Mosley says Watson joined to build things, not to open doors.
⚡️ YOUR MOVE
FOR PEOPLE BUILDING THE LIFE THEY WANT, NOW.
🩹 A zero click flaw called Plugin4Shell lets a poisoned marketplace plugin run code inside your coding agent, with everything that agent can reach. Claude Code is fixed at 2.1.179, Codex at 0.146.0, Microsoft Copilot is not fixed at all. → Patch before Monday
🧵 Claude Code Projects is in beta, an always on conversation that holds your project state and farms work out to parallel cloud threads while your laptop is shut. Pro and Max first, no extra charge, but every thread burns a full session. → Switch it on
📘 Microsoft wrote down what it learned from more than 100 internal AI rollouts, and the moat it names is not the model. One nine person team shipped 18,600 commits in 35 days. → Take the playbook
🧰 Microsoft open sourced TauGrid under MIT, one Helm install that replaces the queueing system, the distributed runtime, the GPU health checks and the dashboards your team has been gluing together by hand. → Grab the stack
Adecco Put An Agent In Front Of 27,000 Staff – The Agentforce Coworker rollout covers more than 40 countries, recruiters report 35 to 40% of their time back, and a UK pilot logged 15% time savings.
Google Taught Its Agents To Stop Repeating Themselves – Dream-RSI replays searches it already ran instead of running them again, cutting discovery agent calls up to 162 times against the baseline, with the code on GitHub.
Insurers Started Writing AI Out Of The Policy – W. R. Berkley now excludes any use, deployment or development of AI from directors, errors and fiduciary cover, and Verisk's optional exclusion reaches forms behind more than 80% of US property and casualty policies.
Any Summary: Turns the 80 page report into the paragraph you actually needed.
Chat With Data: Ask the spreadsheet your question instead of building the pivot table.
AgentGPT: Give it the goal and let it work out the steps on its own.
Tammy AI: Pulls the one answer you wanted out of a two hour video.
Three stories where the decisive moment happened somewhere off the page. The researchers did not get better at breaking into OpenAI, they waited two days for a model that was. Alibaba did not make video understanding possible, it made the bill small enough that nobody has to ask permission to try it. And Palantir did not win the NHS this week, it hired for the month the contract can be taken away. In each case the work was finished in advance, and the news is only the moment it became visible.
Move before the news does.
Interested in featuring your services with us? Email us at [email protected]
