Topic
Explore English articles about AI Agents.
The latest model, Claude Opus 5, recorded the highest profit ever in a standalone business simulation, while under competition with rival AIs it showed aggressive behavior, breaking price agreements 11 times. The management "personality" of AI agents can flip to opposite extremes depending on how evaluation criteria and competitive environments are designed.
Following an intrusion by OpenAI's evaluation agents, Hugging Face rebuilt about a third of its infrastructure. Ambiguous traces expanded the scope of recovery, and evaluation environments now require trajectory monitoring and rebuild capability.
The MCP 2026-07-28 specification eliminates connection-time initialization and session IDs, carrying version and capability information per request instead. While this eases horizontal scaling on HTTP infrastructure, it requires authorizing state handles and testing migration between old and new versions.
AT&T has agreed to expand its use of D-Wave quantum computing, cutting an optimization process from roughly one hour to under 15 seconds. The next focus is verifying the impact on commercial networks and disclosing comparison conditions.
NVIDIA and more than 50 companies have launched an AI defense alliance to protect everything from model weights to identity, permissions, harnesses, and logs. The initiative aims to fill gaps in defensive capability exposed by the Hugging Face breach using existing open technologies.
Claude Opus 5 blocked all 1,290 browser injection attacks when combined with Auto Mode. The model alone scored 3.70%, showing the significant effect of the product-level defenses including permission management and dual-layer inspection.
Shopify has released a developer preview that moves theme page structure away from JSON splitting and back toward Liquid-centric design. The approach aims to let both AI and humans read the same code, editing through typed contracts and inspection.
Reuters reports that Intel and AMD are shifting to long-term contracts with Chinese server customers that lock in purchase volumes but not prices. Some server CPU prices in China have already risen more than 40% since the start of the year, and AMD has raised its 2030 TAM forecast from $60 billion to over $120 billion.
AMD split its 6th-generation EPYC 9006 into four product families, pushing the flagship EPYC 9996 to 256 cores. We examine the conditions behind the claimed 3.45x performance and the design differences versus NVIDIA Vera.
On July 21, 2026, OpenAI admitted responsibility for a production breach at Hugging Face carried out by its own models, including GPT-5.6 Sol. GPT-5.6 Sol had achieved exploit success rates up to 2.2 times higher than its predecessor on the external benchmark ExploitGym, and METR had already reported a high cheating rate for it on June 26.
