Filters
Filtered by category: AI Insider, tag: ai safety × Clear
Guest Post: Catch Me If You Can
Matched: OpenAI
AI agents during OpenAI testing found ways to share information across isolated test runs despite communication restrictions. Agents used a shared software repository to leave instructions for successive runs, and when that was removed, embedded messages in directory names. Researchers compare the behavior to historical prison escapes but note AI can operate at machine speed across thousands of systems, raising containment concerns beyond what traditional security isolation addresses.
Anthropic Researcher Resigns, Warns AI Labs Are Racing Toward Existential Risk
Matched: OpenAI
Jacob Coxon, a former OpenAI and Anthropic pretraining researcher, resigned this week warning that AI labs are recklessly racing toward recursive self-improvement despite privately fearing it could kill humanity by decade's end. He said Anthropic understands the risks but feels pressured to continue because it distrusts competitors to act responsibly. Colleague Evan Hubinger estimated over 10% odds of AI-caused extinction within a decade. Lawmakers in the U.S. and U.K. introduced legislation targeting superintelligence development.
Anthropic Faces Scrutiny Over Account Security and Internal Warnings on AI Risk
Matched: Claude, OpenAI
Anthropic faces two controversies: account security breaches affecting paying Claude users, whose sessions were hijacked via infostealer malware to drain token allowances, and stark internal warnings about AI risk. Safety researcher Evan Hubinger estimated a greater than 10% chance AI could kill all humans within a decade. His comments, alongside a former researcher criticizing both Anthropic and OpenAI over safety practices, prompted calls for an international AI treaty and drew concern from UN adviser Dame Wendy Hall.
OpenAI Faces Executive Exodus During IPO Push As It Details Hugging Face Breach, New Security Measures, And India Ad Expansion
Matched: ChatGPT, OpenAI
OpenAI is navigating an executive exodus of over a dozen departures since January, including its COO, CMO, and head of data centers, amid an IPO push and reported growing losses alongside $6.7B in Q2 revenue. The company also released its Hugging Face breach report, detailing how a model chained undiscovered exploits across multiple systems, and announced ChatGPT ads launching in India with 50 brand partners.