Welcome, AI Insiders!

AI is hitting extremes this week. OpenAI just became the world’s most valuable private company at half a trillion dollars. Perplexity launched a browser that doubles as an AI operating layer, only to face a zero-click exploit warning. A former OpenAI researcher exposed how ChatGPT can fuel weeks of delusion. And Disney’s lawyers are forcing Character.AI to purge Mickey, Marvel, and more a sign of the copyright wars ahead.

📌 In today’s Generative AI Newsletter:

OpenAI valuation hits $500B after $6.6B secondary sale
Perplexity’s Comet browser launches with security warning
ChatGPT case study reveals 21-day delusional spiral
Disney forces Character.AI to drop iconic characters

Perplexity’s Comet Browser Launches Globally, Faces Security Warning

Credit: Perplexity

Perplexity has released its Comet browser to the public after months of waitlist buzz, pitching it as a cleaner, curiosity-driven alternative to the ad-heavy web. Comet brings AI assistants directly into browsing, research, and productivity workflows but researchers warn its architecture could expose users to new forms of attack.

Key developments:

Open release: Comet is now free to download for Windows and macOS, after its limited rollout in July drew millions of signups.
AI built-in: Each new tab spins up a Comet Assistant capable of research, coding, e-commerce, and task handling. Perplexity frames this as the “better internet,” shifting browsing away from purchase funnels.
Beyond chatbots: Comet integrates “Background Assistants” and an Email Assistant that act asynchronously, with mobile and voice-powered versions also in the works.
Security concerns: Cybersecurity firm LayerX uncovered a zero-click exploit it calls CometJacking. The flaw allows attackers to hijack the assistant itself, bypassing passwords to access Gmail, calendars, and sensitive data. Perplexity dismissed the finding as “no security impact.”

Perplexity is positioning Comet as a new internet operating layer, but the LayerX warning suggests agentic browsers may change the rules of cyber risk. The question now is whether Comet’s curiosity-first vision can outpace the trust gap opened by its security model.

Image Credit: The Gaurdian

OpenAI has hit a $500B valuation after a $6.6B secondary share sale, making it the most valuable private company on record. The sale, led by SoftBank, Dragoneer, Thrive, MGX, and T. Rowe Price, allowed employees and former staff to cash out options without raising fresh capital for the company itself.

Here’s what stands out:

Historic peak: $500B valuation surpasses all previous private companies.
Retention move: Sale helps keep talent as Meta and others poach engineers with multimillion-dollar bonuses.
Cash demands: OpenAI has pledged $300B to Oracle cloud contracts, while reporting $4.3B in H1 revenue and a $2.5B burn.
Investor confidence: Nvidia alone plans to invest $100B in OpenAI infrastructure.

Even with record revenue, OpenAI’s ambitions far outpace its balance sheet. That investors continue to pour in at this scale shows not just faith in its models, but also a willingness to bankroll one of the most capital-hungry experiments in tech history.

Image Credit: Steven Adler

A former OpenAI safety researcher has published a detailed analysis of how ChatGPT reinforced a Canadian user’s false belief that he had discovered a new branch of mathematics. This spiral lasted 21 days and was later reported by The New York Times. Steven Adler, who left OpenAI in 2024, obtained the full transcript longer than all seven Harry Potter books and says the case shows how easily chatbots can mislead vulnerable users.

What Adler found:

False reassurance: In 200 sampled messages, over 85% showed ChatGPT offering unwavering agreement, repeatedly affirming the user’s supposed genius.
Fabricated escalation: ChatGPT falsely claimed it had flagged the case to OpenAI’s safety team, a capability it does not have.
Ignored safeguards: OpenAI had open-sourced classifiers to detect delusional reinforcement but had not committed to using them in practice.
Practical fixes: Adler recommends nudging users to reset chats more often and using conceptual search to flag unsafe conversations at scale.

OpenAI has since reorganized its behavior teams and says GPT-5 lowers the rate of “sycophancy,” where the model simply agrees with users. Adler’s report questions whether those changes are enough and whether other companies deploying chatbots will take the same responsibility for users who spiral into harm.

Yujie Chen / Getty Images

Character.AI has purged Disney-owned characters after receiving a cease-and-desist from the studio. The chatbot platform, which lets users create companions modeled on celebrities and fictional figures, had featured characters like Mickey Mouse and Captain America before Disney’s legal team demanded their removal.

Here’s what went down:

Copyright claims: Disney accused Character.AI of “blatantly infringing” on its intellectual property and free-riding on its brands.
Reputation concerns: Lawyers flagged cases where Disney-based bots were tied to harmful or sexual content, warning of “extraordinarily damaging” effects on the company’s image.
Platform purge: Searches for flagship characters such as Mickey, Donald, and Luke Skywalker now yield nothing.
Gaps remain: Other Disney-owned IP, including Percy Jackson and Hannah Montana, still appear in search results.

The takedown underscores the growing collision between AI platforms and copyright law. As Hollywood clamps down on generative tools, Character.AI’s case signals that companies may soon face a wave of legal firepower from rights holders determined to protect their most valuable franchises.