This week in AI: a trillion-dollar equity offer to Washington, the triumphant return of Anthropic's most powerful model, and an AI system rewriting the rules of cybersecurity research.
1. OpenAI Proposes Giving the U.S. Government a 5% Equity Stake
OpenAI is in early talks to voluntarily donate a 5% equity stake to a proposed U.S. sovereign wealth fund — a share worth roughly $42.6 billion at the company's current $852 billion valuation. CEO Sam Altman has championed the idea as a way to share AI's economic gains with the American public, modeled on Alaska's Permanent Fund which distributes resource revenues to citizens as dividends. The proposal, first reported by the Financial Times, would reportedly invite other major AI developers including Anthropic, Google, and Meta to contribute similar stakes. While talks remain preliminary and would likely need Congressional approval for a formal structure, the move signals a deepening entanglement between frontier AI labs and the U.S. government — and a strategic play to reduce political and regulatory pressure on OpenAI's ongoing restructuring.
2. Claude Fable 5 Returns: U.S. Lifts Export Controls, Global Access Restored
Anthropic's Claude Fable 5 — the company's most capable publicly released model — is back online after a two-week forced absence. Originally launched June 9, 2026, Fable 5 was pulled from worldwide access within days after the U.S. Commerce Department imposed emergency export controls citing national security concerns linked to a reported jailbreak vulnerability. Following weeks of negotiations and Anthropic's implementation of updated safety safeguards, the controls were lifted June 30. As of July 1, Fable 5 is available again on Claude.ai, Claude Platform, Claude Code, and Claude Cowork, with a phased rollout to AWS, Google Cloud, and Microsoft Foundry underway. Pro, Max, Team, and select Enterprise subscribers can access the model for up to 50% of their weekly usage limits through July 7, after which access shifts to usage-based credits. The episode marks the first time a commercially released frontier AI model has been subject to — and then returned from — U.S. export controls, setting a significant regulatory precedent.
https://www.anthropic.com/news/redeploying-fable-5
3. AI Finds a 29-Year-Old Bug: Project Glasswing Uncovers "Squidbleed"
In a landmark demonstration of AI-assisted cybersecurity research, Anthropic's Claude Mythos Preview model — deployed under the company's Project Glasswing initiative — helped researchers at Calif.io identify a previously unknown memory-disclosure vulnerability in the widely used Squid proxy server. Dubbed "Squidbleed" and tracked as CVE-2026-47729, the flaw traces back to a parser change from 1997 and allows an attacker to leak cleartext HTTP requests and credentials from other users sharing the same proxy. The bug went undetected for nearly three decades despite Squid's widespread use in enterprise and ISP networks. A fix was released in Squid version 7.6 on June 8, 2026. The finding echoes the original Heartbleed disclosure and has sparked widespread discussion about AI's transformative potential in auditing legacy codebases for buried security flaws at scale.
https://thehackernews.com/2026/06/29-year-old-squid-proxy-bug-squidbleed.html