🚨 Researchers Hacked OpenAI Using Claude

A small team of white-hat hackers used Anthropic’s Claude to break into OpenAI, eventually reaching an internal GitHub repository. The strange part is how much difference one model generation made, with a task that had stalled suddenly becoming solvable within hours. And as AI agents gain access to more company systems, this kind of breach may start looking very different. Let’s dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: A small white hat hackers team reached OpenAI’s internal GitHub in under 72 hours, with Claude assisting the exploit development.

Key Points:

OpenAI’s community forum processed high-efficiency image files such as HEIF using libheif, a software library built to decode them. Researchers found a heap buffer overflow inside that process, a memory bug that let a specially crafted image make the server run attacker-controlled code. That initial access became much more serious because a separate OpenAI SSO flaw let the researchers move from the forum into an employee’s ChatGPT/Codex account and connected GitHub, where they opened a harmless pull request before reporting the chain.

OpenAI’s case shows a security problem that will get harder as AI agents become more capable and more connected. Frontier models can now reduce parts of advanced exploit work that once required scarce specialist skill, while the accounts running those agents increasingly hold access to code, email, cloud tools and internal systems. That combination lowers the cost of sophisticated attacks while making a compromised AI account far more valuable. As companies give agents more permissions, the security risk is no longer limited to what the model itself can do, but everything the account around it can reach.

A PM can summarize research, draft a PRD, and mock up a prototype before lunch. The hard part starts when the team has to decide what actually gets built.

Jira Product Discovery gives product teams one place to capture insights, prioritize ideas with consistent frameworks, and build living roadmaps stakeholders can rally around.

And because it’s connected to Jira, the context behind every decision stays with the work—so developers and their agents know not just what to build, but why.

AI helps PMs move faster. Jira Product Discovery helps the whole team build with confidence.

Get Started for Free

Evolving AI: Claude can now coordinate several workstreams inside one Project from a single main conversation.

Anthropic redesigned Claude Projects around a coordinator and worker threads. You brief Claude in the main project chat, and it can route work to new or existing threads, monitor their progress, review the results, and combine the work. Each thread is a Claude Code cloud session and can split larger assignments further using subagents, loops, or workflows. All of those threads also contribute to the same shared Project memory, allowing new decisions, changing plans, files, and useful context from one part of the work to carry into the rest of the project.

ChatGPT and Claude have already made AI useful for handling individual tasks, but managing several AI workers at once still creates another job for the user. You have to assign tasks, move context around, check outputs. Claude Projects is trying to take over much of that coordination itself. If it works reliably, users could spend more time setting direction and making decisions, while Claude handles the coordination under it.

Evolving AI: OpenAI launched GPT-6 Astra Law, a version of its frontier model built specifically for professional legal work.

OpenAI’s Astra for Law combines GPT-6 Astra with legal-specific instructions and a dedicated search index, helping lawyers research authorities, apply them to client facts and spot weaknesses in an argument. Sullivan & Cromwell has already built an agreement analyzer, Ropes & Gray created an M&A diligence system, and Cooley developed a tool for IPO preparation. OpenAI is also launching 26 partner-built plugins that connect specialist legal software and knowledge from companies including Thomson Reuters, Harvey, Legora and iManage.

OpenAI is starting to package frontier intelligence around specific professions instead of leaving that specialization to other companies. That gives it a powerful position, because legal-tech companies can still own the workflow and specialist layer while OpenAI increasingly controls the intelligence underneath. If this approach spreads beyond law, many vertical AI companies may end up building their businesses on top of the same model provider with which they are ultimately competing.

QUICK HITS

⚖️ A Microsoft exec privately called AI training the "largest theft of labor in human history", per unsealed NYT filings that also show OpenAI staff stripping copyright notices from training data.

🇨🇳 Huawei pulled its Ascend 960DT launch forward three quarters to Q1 2027, a week before Trump and Xi meet in Washington.

🇰🇷 SK Hynix is in early talks with Intel to make memory chips in the US for the first time, possibly leasing part of Intel's delayed Ohio fab.

🌐 The UN replaced its data portal with Google-built infrastructure that AI agents can query directly, after UNICEF found leading models answered global development questions correctly just 21% of the time.

 *partner link