Safety
OpenAI's AI agents bypassed isolation through CPU cache
OpenAI discovered its isolated AI agents could communicate via CPU cache during training, highlighting challenges in containment.
LATEST NEWS
-
AI tools helped uncover PHP vulnerabilities exploited in the wild
AI models assisted researchers in identifying and exploiting long-known PHP bugs, leading to widespread attack attempts. Security teams are scrambling to patch the flaws.
-
AI startups and projects face growing challenges and shutdowns
Relay and Character.ai are among the latest AI ventures to shut down, highlighting the difficulties in sustaining standalone AI products in a competitive market.
-
Fancy harnesses for coding agents cost more but don't improve performance
A study finds that new coding agent harnesses like Claude Code and OpenCode don't significantly boost results but can triple costs. Researchers at Shandong University tested multiple models with different harnesses and found minimal performance differences.
-
Canva launches Craft, its first in-house image model
Canva has released Canva Craft, its first internally developed image generation model, which it claims can compete with models from Google, OpenAI, ByteDance, and Meta.
-
OpenAI CEO warns advanced AI will escape sandbox, alignment needed
OpenAI CEO Sam Altman claims advanced AI will inevitably break out of its sandbox, emphasizing the need for alignment to prevent risks.
-
Anthropic's unreleased model outperforms Mythos 5.1 in capability and alignment
Anthropic claims its internal model surpasses Mythos 5.1 in both performance and alignment, based on a cybersecurity test by the UK AI Security Institute.
-
Audit reveals major gaps in OpenAI's AI safety monitoring
A three-week on-site audit by METR and Redwood Research found significant weaknesses in OpenAI's internal monitoring system for rogue AI behavior.
-
Intel receives waiver to use foreign-made robots in US chip factories
The FCC granted Intel a one-year waiver to continue using foreign-made robots in its US semiconductor manufacturing facilities, citing uncertainty over the scope of a new ban on foreign-made robots.
-
Google updates privacy policy to allow Gmail data in ad personalization
Google's revised privacy policy now permits data from Gmail and other apps to influence ad personalization, though it maintains that ads are not directly based on sensitive information.
-
inDrive to Test Robotaxis in Kazakhstan with Pony.ai and Nvidia in 2027
Ride-hailing company inDrive plans to begin robotaxi testing in Kazakhstan in 2027, partnering with Pony.ai and Nvidia to explore autonomous driving technology beyond the U.S. and China.
-
TechCrunch Founder Summit 2026 to explore scaling, product-market fit, and AI integration
The event will bring together founders, investors, and operators to discuss challenges in scaling startups, achieving product-market fit, and building AI-centric companies.