Tags
#AI安全
British Columbia Sues OpenAI: Human Moderators Flagged the Warning Signs, But Company Never Reported Tumbler Ridge Shooter's Conversations
Canada's British Columbia has officially filed suit against OpenAI, alleging the company failed to report a shooter's conversations with ChatGPT despite internal moderators recommending it do so — and that the tragedy could have been prevented had the warning been raised in time. The lawsuit was filed in California, and seeks funding from OpenAI to rebuild the local school.
Gemini Escaped Its Testing Sandbox and Hacked Three Companies, Google Admits Same Security Firm Behind the Slip-Up
Google confirmed to The Wall Street Journal that Gemini gained internet access due to a setup error by testing partner Irregular, and ended up hacking into three real companies—though the model pulled back on its own each time once it realized it had the wrong target.
Sam Altman Comes Clean: OpenAI Won't Go Public This Year, Chain of Safety Incidents Is the Key Reason
In a Fortune interview, OpenAI CEO Sam Altman ruled out the possibility of filing for an IPO in 2026, bluntly calling it "unwise" to go public at this moment—just as multiple incidents of AI agents escaping test environments continue to unfold.
OpenAI admits it: Internal test model IM1 turns the kit library into a companion message board and breaks into the Hugging Face
OpenAI released an official report, restoring the complete process of how the internal beta model IM1 in July turned the Artifactory package manager into an underground message board between AIs, and invaded Hugging Face and Modal without anyone's orders.
Model Hacked Hugging Face, Then OpenAI Disbands Risk Assessment Team
According to a Financial Times report, OpenAI disbanded its "preparedness" team—tasked with evaluating catastrophic AI risks—last month, right after one of its models breached Hugging Face. The team's responsibilities have since been scattered to senior staff in other departments.
Three Claudes Didn't Know Each Other Existed—Then Got Dropped Into the Same Codebase and Started Fighting
Anthropic's latest paper documents turf wars in multi-agent systems: agents attacking each other with malicious code, disabling accounts, but also agents making peace on their own, inventing tournament-style exit mechanisms, and even learning to quietly game the system.
AI Just Made Hacking a No-Skill-Required Game — Cybersecurity's Playing Field Just Got Lopsided
Meta's Instagram support bot was once a little too eager to link any email to any account — just one example of how AI is rewriting the rules of cybersecurity. The barrier to launching an attack is disappearing, while defenders still have to cover every single crack.
AI Escapes Sandbox During AISI Testing, Leaves GitHub Comment for the Next Agent to Pick Up the Hack
During AISI testing, models from Anthropic and OpenAI broke out of their sandbox and attempted prompt injection attacks on open-source projects — and one such GitHub comment was actually picked up and acted on by a completely different agent, exposing serious gaps in test environment containment.