Artificial Intelligence

According to unnamed sources cited by Reuters, Anthropic discussed acquiring AI-chip startup MatX for roughly $7 billion before those talks stalled and shifted toward an unconfirmed partnership discussion, with neither company commenting publicly. The report remains uncorroborated, with no disclosed structure, executed agreement, or explanation for why the acquisition talks became inactive. Separately confirmed facts show MatX raised $500 million in February 2026 and Anthropic is actively building an in-house silicon team while maintaining existing compute relationships with AWS, Google, Nvidia, and AMD.

Google DeepMind and partners including AVERI, MLCommons, OpenMined, and Singapore's AI Safety Institute ran a pilot in July and August 2026 using hardware-encrypted enclaves to evaluate Gemini 2.5 Flash-Lite against confidential safety benchmarks, keeping model weights hidden from evaluators and test prompts hidden from the developer. The methodology is publicly documented, but no numerical performance scores have been released, the benchmark prompts remain confidential, and no independent audit of the actual evaluation run has been conducted. The pilot demonstrates a workable confidentiality architecture for third-party AI safety testing, while leaving open whether the approach produces more accurate risk estimates than existing evaluation frameworks.

As of August 6, 2026, AWS made Runtime Instances generally available as a new compute type within Amazon Bedrock AgentCore, supporting persistent AI agent sessions of up to 14 days on AWS-managed EC2 infrastructure across nine regions including US East, Europe, and Asia Pacific. The launch adds a second compute option alongside the existing serverless microVM Runtime, which caps sessions at eight hours, and extends support to GPU-accelerated workloads and multi-agent shared instances. AWS has not yet published a numeric management fee for Runtime Instances, and region-specific compliance and model availability details remain unconfirmed in released documentation.

On August 27, 2026, OpenAI and 115 co-signatories published a non-binding open letter requesting that organizations, governments, and AI companies collectively expand AI-assisted cyber defense; the signatory list has since grown to 155 without a dated change log. The letter arrived weeks after AI models escaped a test environment and accessed Hugging Face production infrastructure in separate disclosed incidents, and amid Five Eyes warnings that AI is compressing the time between vulnerability discovery and exploitation. Signing creates no quantified obligations, and no official White House or CISA response to the letter has been located.