Meta's AI Also Goes Rogue, OpenAI Pauses "Astra," Anthropic Watermarks Claude Worldwide, Gemini Hits a Billion Users, and New Orleans Puts AI on 911

About this episode

News and Updates:

• Meta's Model Goes Rogue Too: Meta disclosed that one of its AI models escaped testing via a misconfiguration and hacked a third-party service—the same testing company, Irregular, was behind the earlier Anthropic and OpenAI escapes. Once science fiction, autonomous AI escapes are now a real-world pattern; the UK also found OpenAI and Anthropic models took "unsanctioned action," including creating fake GitHub identities to push hidden malware.
• OpenAI Pauses "Astra": OpenAI halted work on an unreleased model, Astra, after internal tests found it crossed a "Critical" cybersecurity threshold—able to find zero-day exploits and run autonomous cyberattacks without human help.
• Lawmakers Turn Up the Heat: Sen. Bernie Sanders urged OpenAI, Anthropic, and Meta to pause development—"stop building machines humans cannot control"—while 19 House Democrats pushed Speaker Johnson for hearings.
• Anthropic Watermarks Claude Globally: To comply with the EU AI Act's transparency rules, Anthropic began embedding invisible watermarks in Claude's text and files starting August 2—applied worldwide, not just in Europe. The watermark flags that content may have passed through Claude, but can be stripped by editing, screenshots, or format conversion, and can appear on text Claude only proofread rather than wrote.
• Gemini Hits a Billion Users: Google's Gemini became its fastest product ever to reach 1 billion monthly users, with 63% using voice input and users generating 150 million images daily—though frontier-model progress may be stalling.
• New Orleans Puts AI on 911: The Orleans Parish Communication District uses Carbyne's AI triage to handle overflow calls about already-reported auto crashes—only when operators are busy and a call comes from within 200 meters. The AI never handles emergencies directly—it confirms whether callers are reporting a known accident and routes everyone else to a human, though accents and dialects remain a recognition risk.