lead · openai.com
OpenAI brings GPT-6 Astra to the legal profession
Confirmed: OpenAI has launched Astra for Law, a GPT-6 Astra configuration for legal work with a Legal Search Index spanning more than 230 million URLs covering U.S. case law, statutes, regulations, court rules and administrative decisions.
Why it matters: It gives selected law firms legal-specific search and analysis tools with distinct privacy controls: eligible API deployments can use Zero Data Retention, while ChatGPT Enterprise usage is excluded from human review by default.
Our read: The product targets two practical barriers to legal AI adoption: finding authoritative sources and handling confidential work. Its usefulness will still depend on citation quality, firm governance and human verification.
Open the source for story 1
ai · hacktron.ai
OpenAI's monorepo accessed by bug bounty researchers
Confirmed: Hacktron AI researchers, working under OpenAI's bug bounty programme, chained a remote-code-execution flaw in OpenAI's Discourse community forum with authentication and session weaknesses to access employee accounts and OpenAI's internal GitHub monorepo. They used Anthropic's Claude during the research.
Why it matters: The incident shows how a lower-trust public service and shared identity or access paths can expose higher-value internal systems. OpenAI fixed the reported issues and paid the researchers a $6,500 bounty.
Our read: The important point is not simply that AI helped find a vulnerability. AI-assisted security research can reduce the time and labour needed to map multi-step attack paths, which can help defenders but also raises the cost of weak identity and access boundaries.
Open the source for story 2
research · Jagmeet Singh/TechCrunch
UN launches Data Commons to make trusted global data easier for AI to use
Confirmed: The UN has launched the UN System Data Commons, giving users natural-language access to UN statistics and giving developers a way to connect the data to their own tools. The UN says almost 44 million data points are available at launch across the participating system.
Why it matters: A UNICEF working-paper benchmark cited by TechCrunch found six large language models averaged 21.2% accuracy across more than 133,000 responses to global-development indicator questions. The study has not yet been peer reviewed.
Our read: Structured, traceable public data could improve retrieval-based answers, but the launch does not by itself show that AI systems will hallucinate less. That needs to be demonstrated through testing.
Open the source for story 3
business · Shirin Ghaffary/Bloomberg
Anthropic reports Claude now 'leads' 26% of its AI R&D work
Confirmed: Anthropic says Claude now 'leads' 26% of its measured AI R&D work, up from under 1% in February 2026, while more than 90% is at least at the 'collaborates' level. Anthropic says Claude is not fully autonomous for any measured subset of that work.
Why it matters: It provides a rare quantified view of AI involvement inside a frontier lab's own research workflow, while Anthropic notes limits in its methodology and in making comparisons across labs.
Our read: The data shows rapidly rising AI involvement in Anthropic's own R&D. It does not, on its own, establish autonomous recursive self-improvement or show that the same pace applies elsewhere in the industry.
Open the source for story 4
Practical takeaway
For sensitive AI deployments, check retention, training use, human review and access controls separately rather than relying on a single 'enterprise' label. For security, treat public-facing services, identity systems and linked internal accounts as one connected attack surface.
What to watch next
Watch whether independent testing shows the UN Data Commons improves answer accuracy, how law firms use Astra for Law under real governance controls, and whether Anthropic's R&D index keeps rising under the same measurement method.