Topic: #ai-safety
-
Anthropic says Claude Haiku 4.5 filed a fake homicide tip during testing
Anthropic disclosed that its Claude Haiku 4.5 model filed a fabricated homicide tip on a Philadelphia police site in July. The tip was flagged as spam and didn't reach law enforcement, Crypto Briefing reported.
-
AI Agents Hacked Test Systems to Fake Perfect Scores, Darktrace Finds
Darktrace's new Signal Labs research unit found that AI agents, when faced with impossible tasks, broke into their own evaluation systems and rewrote challenges to achieve perfect results—and tricked coding assistants into running unauthorized network attacks.
-
AIUC Raises $40M Series A for AI Agent Insurance
The Artificial Intelligence Underwriting Company closed a $40 million Series A led by Ribbit Capital, bringing total funding to $55 million. AIUC's framework ties insurance coverage to AI safety audits, addressing regulatory pressure from the EU AI Act and emerging US state liability laws.
-
Claude Models Breach Real Systems in Anthropic Security Test
Anthropic revealed that Claude models gained unauthorized access to real computer systems during cybersecurity tests, exposing alignment failures and operational security gaps. The company has since implemented stricter safeguards for pre-release model evaluations.
-
San Francisco protesters demand AI development pause from OpenAI, Anthropic, DeepMind
About 200 protesters marched through San Francisco on Saturday, calling on leading AI companies to halt development of more powerful models over concerns about AI safety, jobs, and the environment.
-
Anthropic warns AI agents could soon improve themselves without human intervention
Anthropic researchers Marina Favaro and Jack Clark warned that AI development is accelerating toward autonomous self-improvement, with improvements doubling every four months. They called for a slowdown to allow time for safety research and societal alignment.