Good morning. Mark Zuckerberg’s chief advisor in Washington is set to leave the company after 15 years of service.
Why did global affairs head Joel Kaplan leave one month before the U.S. midterm elections? Neither he nor the company has provided detailed explanations. Perhaps this former advisor to George W. Bush, who helped Meta build ties with the Trump administration, wished to depart before widespread predictions anticipated Republican losses. Perhaps he left because Meta’s president, Dina Powell McCormick—a former Trump advisor whom Meta hired just nine months ago—might be a better representative for government communications. Or perhaps he left because Zuckerberg believed Muse could handle the role.
I was joking. By the way, I asked Muse whether AI could replace policy advisors. Meta’s personal AI agent replied: “It can partially replace, but not fully replace—and that distinction matters.” The most likely outcome is that this role will be “less about replacement and more about reshaping.” As for replacing CEOs? Muse said: “Even less likely.” Smart robot. Here’s today’s tech news. — Andrew Nusca
Have an idea? A suggestion? Something explosive? Feel free to reach out.
Breaking News
Gray Swan revamped for AI agent red team competition
AI security startup Gray Swan helps leading AI companies, including Anthropic, OpenAI, and Google DeepMind, conduct stress tests on their models before release. The company is overhauling its public red teaming platform to remain relevant in the age of AI agents.
The Pittsburgh-based company launched its redesigned Arena on Tuesday. It is an online platform where participants compete for prizes by attempting to lure AI models into violating their safety safeguards.
In conjunction with this relaunch—termed Arena 2.0 by Gray Swan CEO Matt Fredrikson—the company has launched its largest competition to date: a challenge called "Indirect Prompt Injection (IPI) Black," with a total prize pool of $300,000, sponsored by Anthropic, Meta, Google DeepMind, and Reflection AI. New attack scenarios will be released weekly until November 2.
In indirect prompt injection attacks, attackers hide instructions within content that AI agents may encounter while performing tasks, such as web pages or forum posts, aiming to hijack the agent. As companies allow AI agents to browse the web and act on users' behalf, this has become a major security risk. This red team testing platform includes a sandbox environment where these tests can be safely executed without causing real-world harm.
Fredrikson told Fortune that anything made public on Arena corresponds to the model’s “guarded version.” (Gray Swan also operates a private Arena for leading AI labs, sometimes using unguarded models with participants who have been vetted and signed nondisclosure agreements.) For all tests, the company builds realistic simulated environments and conducts red teaming assessments both internally and through third parties to determine whether these environments could be escaped, while monitoring agent behavior.
Fredrikson said, "Have we seen any indications that it may have discovered a zero-day vulnerability and already breached the sandbox environment? We can monitor for this."
Fredrikson said that when Gray Swan launched the initial version of Red Team Arena in September 2024, most AI applications were still chatbots, and models performed poorly at adhering to their own guidelines. He told Fortune that during the first competition, participants could easily jailbreak models from all major AI providers.
Since then, AI labs have significantly improved model robustness, and the systems deployed by enterprises have also evolved. Today, the technical expertise required to configure these test environments and set up attacks has increased substantially.
Fredrikson said, “It’s no longer about chatbots. There may still be chat-like interfaces, but what’s really running are these systems, these agents.” — Jeremy Kahn
Mystery Box
They will have to burn all their furniture.
Click here to read the full article.
SMS
- DeepSeek stated that the ban by a U.S. federal agency is an "overextension of the concept of national security."
- Apple has added a new U.S. titanium supplier for its new iPhone Duo.
- Several banks in London experienced hours of downtime due to a Google Cloud outage caused by a "power event."
- OpenAI acquired Orchestra Bio, a company that develops AI agents for managing drug research and development.
- Daytona has reached a $1 billion valuation. This U.S. startup is building an "agent cloud."
- Mural raised $15 million. This startup, founded by former Palantir engineers, aims to help banks connect to the emerging robotics economy.
- Google allows AI applications to run on TPUs and Google Cloud without requiring code rewrites.
Final blow
Before Anthropic's planned IPO, its S-1 filing with the U.S. Securities and Exchange Commission has been leaked. (The company submitted the documents confidentially in June.) Three points are noteworthy:
- Anthropic lost over $80 billion in the first half of this year.
- Anthropic's spending on computing power over the next few years could reach $750 billion.
- Of Anthropic's $16.3 billion in revenue in the first half of 2026, one-fifth came from just two customers (who? According to previous reports, they may be Amazon and Google).