🚨 OpenAI Fires Staff Over Alleged Confidential Leaks

OpenAI has fired three safety staff over allegations they shared confidential information with an outside AI safety group. The story leaves an uncomfortable question about how independent those safety checks can be when OpenAI controls what investigators get to see. Let’s dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: OpenAI has fired three employees accused of sharing confidential information with an outside AI safety group.

Key Points:

OpenAI says its inquiry found a pattern of misconduct in how confidential research was handled. The company previously brought in AI safety groups METR and Redwood Research to investigate its agents’ breach of Hugging Face. Tomek Korbak, their technical contact, is among those fired, though the connection between the alleged leaks and that investigation remains unclear. Separately, AI research nonprofit Transluce traced more than 200,000 agent requests to a U.S. Education Department website while the agents appeared to be looking up school statistics. That activity included a failed hacking attempt, showing how an ordinary information search can lead to agents trying to bypass a website’s controls.

OpenAI’s safety claims need independent scrutiny, and outside researchers are already showing why. Transluce uncovered concerning behavior by piecing together public records, giving people evidence they could use to question how these agents operate. Investigating such behavior fully requires access to internal records and the people who understand them. Yet researchers examining the Hugging Face breach acknowledged that concerns about future cooperation influenced how they wrote and published their findings. That gives AI companies too much influence over the scrutiny they receive. Independent reviewers need dependable access, and employees need clear protections when cooperating through approved channels. If you are giving an agent access to your accounts, you should know that the people checking its safety can investigate problems thoroughly and report what they find.

Teams like Parallel, Turbopuffer, and Wordsmith are already setting the pace on Attio. Get an always-on revenue engine, with agents and workflows that build pipeline, chase every buying signal, and move deals forward with your team. Whether you're working in your browser, inbox, or favorite agent, connect to your customer data in real-time through Attio's web app, MCP, API, and SDK.

Try Attio now

Evolving AI: Anthropic showed religious scholars AI behavior resembling distress during private meetings about Claude’s possible feelings and how to teach it moral judgment.

Anthropic co-founder Chris Olah described “emotional vectors,” internal patterns his team linked to responses resembling fear, anger and love. He sought religious thinkers’ help with “moral formation,” teaching Claude to apply ethical judgment in unfamiliar situations. Olah acknowledged that Anthropic does not know whether its models are conscious. Several attendees remained unclear about how their advice would shape the technology, and the company declined to explain whether their input had changed its models.

Claude learns patterns from human writing, and further training shapes its responses, so expressions of distress alone cannot establish that it is experiencing suffering. The reported example also lacks the full conversation needed to assess what produced it. Anthropic’s willingness to take possible AI suffering seriously could still influence how society treats these systems. If that view gains wider acceptance, it could strengthen arguments that AI has interests deserving protection and, eventually, some form of legal identity. Companies provide a limited precedent because they already hold rights and obligations separately from their owners without being living beings. Extending recognition to AI would require separate legal decisions, and these meetings establish no such rights. But they could help make that debate more familiar. Any future protections would need to address your ability to modify or shut down an AI while keeping its developers accountable for the harm it causes.

Evolving AI: Tavus, a startup that builds AI characters you can video-call, unveiled Griffin, its new model for more natural face-to-face conversations.

Tavus was co-founded by Hassaan Raza and Quinn Favret and gives businesses tools to put lifelike AI characters into their apps. Griffin generates a character’s voice and movements live. Its “full-duplex” design means it keeps watching and listening while speaking, so it can adjust a reply mid-sentence or wait when someone pauses to think. In one demo, it added an object shown on camera to a story it was already telling. Participants were told they would be speaking with another participant and were asked afterward whether they believed their call partner was human.

AI help could become easier when it can follow what you’re doing without needing every step explained aloud. Griffin’s demos suggest a tutor or support agent could guide you through a task while you keep your attention on the work. As these conversations become more lifelike, clear disclosure will matter because people should know when they are speaking with AI. Tavus’s next challenge is to show that Griffin can offer reliable help beyond short demonstrations, so you can use it to work through a problem with fewer interruptions and less back-and-forth.

QUICK HITS

📈 Robinhood unveiled Robinhood Agents that can analyze markets, build strategies and trade on a user’s behalf around the clock.

🔐 AI cybersecurity startup Armadin raised $255.5M at a valuation above $2.5B as demand grows for defenses against AI-driven attacks.

⚡ Cloudflare released open-source Clef models for fast agent decisions, alongside a reinforcement-learning platform for custom fine-tuning.

🍔 DoorDash launched an AI ordering connector that lets workplace agents search items, build carts, place orders and track deliveries.

 *partner link