Tuesday, 28 July 2026

Cyera signed a letter of intent to acquire Oasis Security for about $1 billion, mostly cash, adding non-human identity security to its data-security platform.

Cursor launched a ₹649/month India-only tier built on its own models, undercutting its $20 Pro plan as it courts its third-largest, fastest-growing market.

Gemini API Managed Agents adds pre/post tool-call hooks, a max_total_tokens budget cap, cron-style scheduling, and a free tier, defaulting to Gemini 3.6 Flash.

Anthropic says Claude Mythos Preview cut a post-quantum signature scheme's effective keysize in half and sped up a reduced-round AES attack up to 800x.

LFM2.5-Encoder-230M and -350M add bidirectional text understanding to Liquid AI's edge model family, built for classification and search on a CPU.

Cognizant joined Anthropic's Global Premier Partner tier, training 30,000+ staff on Claude and citing a 40% cut in contract-review time for one client.

Dario Amodei says Anthropic never asked for a blanket ban on open models, and instead wants chip controls, distillation limits and mandatory safety tests.

The 2026-07-28 Model Context Protocol spec is out: no more session handshake, plus caching, extensions and a formal deprecation policy.

A new audit method called HackDetect found exploitable shortcuts in 67% of agent benchmark traces reviewed, inflating reported capability scores.

OpenAI's CEO said on a podcast that AI has entered a self-improving acceleration phase; experts note there's no agreed scientific test for that claim.

OpenAI's analysis of 800,000 US ChatGPT messages found workers routinely using it for tasks outside their own occupation, a pattern it calls task crossover.

AWS's new AgentCore tutorial combines LangGraph orchestration with Strands agent reasoning for a market-surveillance system with checkpoint-based recovery.

NVIDIA's Jetson Orin Nano Super, 67 TOPS in a compact dev kit, powers demos from an autonomous toy car to an on-device voice-and-vision robot assistant.

Ai2 says a North America wildfire-risk run used 19,600 CPUs and 994 GPUs to hit a 155x speedup, cutting 4,737 serial compute hours to 30.5 wall-clock hours.