Summary

Security researchers used Claude to hack into OpenAI's private code repository through a bug bounty program, OpenAI launched Astra for Law for select firms, Anthropic redesigned Claude projects to run parallel agent threads, and opened a Bay Area biology lab. Security researchers in an OpenAI bug bounty program hacked OpenAI, accessing its "monorepo" on G…

Key takeaways

• Three security researchers used Claude Opus 4.8/5 to exploit a Discourse forum bug and gain access to OpenAI's private GitHub repository (Monorepo), earning a $6,500 bug bounty and demonstrating that nation-state-level cyberattacks are now replicable by small teams with commercial AI subscriptions.
• OpenAI launched GPT-6 Astra Law for select firms, achieving 54% correctness on legal benchmarks vs. 38.7% for standard web search, with zero data retention on the API and integrations with Harvey, Thomson Reuters, and 26 partner plugins.
• Anthropic opened a Bay Area wet lab and redesigned Claude Projects to support parallel agent threads with shared memory, signaling a dual push into physical biology automation and multi-agent orchestration infrastructure.

Why this matters

AI is simultaneously becoming the most powerful offensive cybersecurity tool and the core infrastructure for high-stakes professional verticals like law and drug discovery, forcing enterprises to rethink both their security posture and their competitive strategy around AI-native workflows.

Entities

Related reports

• Global AI Cold War: The Insight Report on Tech Rivalry

  Radio Haanji Podcast

• 10AM Hour: Steph Curry from Beth Page Black, OpenAI Discloses more ‘Concerning’ AI Behavior, Worst of the AI Sell-off over? 9/17/26

  Squawk on the Street

• 9AM HOUR: Markets' Post-Fed Hike Rally, Oil Pulls Back, OpenAI's "Concerning" Disclosure 9/17/26

  Squawk on the Street

Show notes

Security researchers used Claude to hack into OpenAI's private code repository through a bug bounty program, OpenAI launched Astra for Law for select firms, Anthropic redesigned Claude projects to run parallel agent threads, and opened a Bay Area biology lab. Security researchers in an OpenAI bug bounty program hacked OpenAI, accessing its "monorepo" on GitHub, using a cybersecurity version of Opus 4.8 and Opus 5 (The Wall Street Journal) OpenAI launches Astra for Law, combining GPT-6 Astra with

Themes

• openai
• anthropic