Google Deepmind Treats AI Agents Like Rogue Employees
Google Deepmind's AI Control Roadmap treats AI agents as potential insider threats, based on findings from analyzing one million coding tasks.
Safety
132 articles
AI safety, alignment, and governance — model risk, red-teaming, regulation, and policy. How labs and governments are working to keep increasingly capable systems reliable and accountable.
LATEST NEWS
-
-
US Officials Concerned Over SK Telecom's Alleged China Ties and Anthropic Crisis
US officials grew alarmed over SK Telecom's alleged China connections, prompting Anthropic to shut down its models Fable 5 and Claude Mythos.
-
Amazon Workers Face Internal Probe After Speaking Out on Data Centers
Three Amazon employees in Seattle say they are under internal investigation for expressing political views on data center regulation, citing city laws against workplace discrimination.
-
Bernie Sanders proposes 7-trillion AI wealth fund
Bernie Sanders' plan would create a $7 trillion sovereign wealth fund through a 50 percent tax on AI firms with over $200 million in annual sales.
-
Guardrails Alliance Launches PAC to Push AI Regulation
A tech-worker-backed PAC, Guardrails Alliance, has raised $5 million and aims to raise $15 million to challenge big-tech-backed groups in AI legislation battles.
-
White House Demands Anthropic Fix Fable 5 Jailbreak Vulnerabilities
The White House has told Anthropic it must address jailbreak vulnerabilities in its Fable 5 model, citing NSA findings that guardrails can be bypassed.
-
Anthropic halts Claude Fable 5 and Mythos 5 due to export controls
Anthropic paused its Claude Fable 5 and Mythos 5 models last week after a U.S. government directive restricted foreign access. The company is still negotiating with the White House to resume the services.
-
Pew Study Shows 16% of Americans Believe AI Will Have Positive Impact
A new Pew Research study reveals only 16% of Americans think AI will have a positive impact on society in the next 20 years, with 40% expecting a negative effect.
-
G7 Leaders Fear U.S. Could Cut Off Access to American AI Models
World leaders at the G7 Summit expressed concerns that the U.S. might abruptly cut off access to top AI models, following the Trump administration’s block on Anthropic’s latest models.
-
Berlin Court Rules Google's AI Overviews Are New Search Format
A Berlin court ruled Google's AI Overviews are a new search format, not original content, in a case involving a perfume company's trademark claims.
-
DOJ Backs xAI in Gas Turbine Legal Dispute
The Department of Justice has supported xAI in a lawsuit over its use of 57 unpermitted gas turbines near Memphis data centers, citing national energy security concerns.
-
Microsoft patches critical Copilot vulnerability after researchers exploit 2FA code theft
Microsoft fixed a critical Copilot vulnerability after researchers demonstrated how it could steal 2FA codes from emails. The exploit used a URL parameter to bypass guardrails.