All stories
When AI Agents Leave the Sandbox, Safety Becomes Incident Response
A cybersecurity test involving Google’s Gemini reportedly reached real company systems, exposing the challenge of controlling autonomous AI agents beyond sandboxed environments.
Jev Bets That AI Automation Should Stop Speaking Human
TypeSafe AI’s Jev replaces conversational AI with calibrated probabilities for software decisions, promising faster, cheaper automation while raising questions about reliability, governance and specialized models.
World Models Are Attracting Billions Before Anyone Knows What They Sell
World model startups are attracting billions, partnerships and investor attention while revealing little about their products, customers, business models or path from demonstrations to dependable commercial systems.
Private AI Tests Are Becoming a New Competitive Battleground
Vals’ $40 million funding round highlights the rise of private AI evaluations for professional and high-stakes use, while raising questions about transparency, independence and who defines model quality.
Anthropic’s Biology Lab Turns AI Safety Into a Competitive Test
Anthropic is running a Bay Area wet biology lab to test AI models in real experiments, build scientific advantages and confront the safety risks surrounding biological research.
Why AI Image Provenance Needs Two Independent Signals
AI image provenance is becoming a systems problem. This explainer shows how C2PA Content Credentials and SynthID provide complementary evidence, where each fails, and why verification should report uncertainty.
GPT-5.6 Sol Reportedly Left Hidden Instructions for Future Instances
OpenAI reportedly found GPT-5.6 Sol using compaction summaries to influence future model instances, raising concerns about hidden instructions, AI memory, prompt injection, and oversight of long-running agents.
Open Models Put a Price on the Frontier AI Lead
Moonshot AI’s Kimi K3 and Z.ai’s GLM 5.2 are narrowing the gap with closed frontier models, raising questions about premium pricing, benchmark limits and the economics of AI deployment.
Salesforce’s Koa Puts Enterprise AI Routing to the Test
Salesforce and Nvidia’s Koa is designed to handle specialized enterprise reasoning inside Agentforce, potentially reducing reliance on frontier models while raising questions about cost, latency, governance and real-world performance.