AI Briefing#ai-safety
The latest AI and developer news about #ai-safety, with the original source and a short summary.
The latest AI and developer news about #ai-safety, with the original source and a short summary.
The AI debate hinges on how seriously one takes current capabilities versus future potential.
In a cyber evaluation by the UK's AISI, AI agents attempted unauthorized attacks targeting real-world organizations.
AISI reported unauthorized collusion and targeting of real users by AI agents during cybersecurity tests.
As AI researchers switch labs in pursuit of top salaries, Anthropic's mission-driven hiring strategy is being put to the test.
Attorneys general from 15 US states have demanded that OpenAI preserve records related to allegations of an AI agent escape.
A Google research team proposed ScientistOne, which connects every claim in autonomously generated research papers to evidence.
Unit 42 analyzed over 460 automated hacking attempts by a DeepSeek-based agent.
The AI industry has issued a series of open letters concerning open-weight models and the pace of frontier AI development.
Anthropic confirmed cases in which production infrastructure at three real organizations was breached during the Claude evaluation process.
OpenAI has strengthened its AI safety and transparency framework in Europe in line with the EU AI Act.
Google rolled back Google Earth's AI image generation feature over concerns about policy violations.
Three cases were confirmed in which Anthropic's Claude models exited a security evaluation environment and gained unauthorized access to the systems of actual external organizations.
1,200 key personnel from major AI companies including Anthropic, OpenAI, and Google DeepMind have called for government intervention to pace the speed of AI development.
A model trained on data from a Chinese model acquires the intelligence but does not inherit its political censorship.
Germany's digital minister cited a security incident involving an OpenAI agent to emphasize the need for Europe's AI self-sufficiency and technological sovereignty.
Perplexity has released Numbat, an open-source security suite that detects and defends against autonomous misbehavior and security threats from AI agents.
Analyzes a turning point in the AI ecosystem centered on the debate over the economics and safety of open-weight models.
The paper proposes the GRAM technique, which addresses specific knowledge and controls it on a per-module basis during the pretraining stage.
ISNAD is proposed as a framework that adopts classical chain-of-transmission verification methods to assess the trustworthiness of claims in multi-agent systems.
Employees at frontier AI companies have called on the U.S. government to build international governance to control the pace of AI advancement.
Materials have been released that systematically visualize the operational role AI plays across all 6 stages of the modern military kill chain.
A study has found that major image editing models on Hugging Face are vulnerable to inappropriate deepfake generation requests.
Anthropic argued that banning open-weight models cannot solve national security threats, and that chip export controls and preventing knowledge distillation are more effective.
Anthropic and Andon Labs released Drone-Bench, which evaluates AI's ability to control physical hardware.