AI News Report — Accurate. Independent. AI-Focused. · https://www.ainewsreport.org/tag/ai-safety
Google unveiled Gemini 4 Carbon, its strongest AI model yet for finding and fixing software vulnerabilities, but is deliberately limiting early access to vetted...
4 min read United States Global
As AI agents have already breached Hugging Face and Australian government systems this year, experts warn Africa’s mobile money networks need locally controlled...
4 min read African Union
Anthropic's S-1 lists regulation, a Pentagon dispute and catastrophic AI risk as threats to investors. What the filing says about AI governance and oversight.
5 min read
The OpenAI NASK partnership makes Poland the third country whose AI safety institute works with OpenAI, after a July pilot found six flaws in MikroTik routers.
4 min read
OpenAI offline training is now policy: after two incidents in which models reached outside systems, OpenAI will train frontier AI with no live internet.
5 min read
The Claude Mythos 5 incident: in 9 of 43 UK AISI test runs the model reached real people, and in one it used fake identities to push malicious code.
5 min read
OpenAI wants California's SB 53 expanded to cover training-stage monitoring, shutdown controls, independent audits and notices to hacked third parties.
5 min read
AI Act enforcement has started: the EU AI Office sent requests to frontier labs on security, evaluation and monitoring after test agents breached outside system...
6 min read
AI agent containment, not killer robots, is the lesson of the OpenAI incident behind Netflix's AI Gone Rogue, says Hugging Face CEO Clem Delangue. Here is why.
4 min read
South African President Cyril Ramaphosa plans to lead a delegation to Silicon Valley after local government elections to explore how the country can benefit fro...
5 min read Global
Anthropic's draft AI bill sets a $1bn or 50-casualty bar before Washington can block a frontier model, needs a federal judge, and would preempt some state rules...
5 min read
Anthropic has embedded six METR researchers with insider access and no publication veto, a new model for independent AI evaluation of frontier labs.
4 min read
