Anthropic split its Cyber Verification Program into three access tiers that relax its cybersecurity blocks step by step for vetted security teams, SiliconANGLE reported. It also folded in Project Glasswing, which since April has given organizations that secure critical software access to its Claude Mythos model. Every tier offers Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1. The entry Defense Access tier covers incident response and malware analysis for teams at companies, universities and government bodies, and Red Team Access permits authorized penetration testing. Specialized Access, open only to organizations cleared to test systems such as power grids and telecom networks, requires in-depth vetting with the U.S. government. In Anthropic's tests, 46 of 50 offensive benchmark runs were stopped at some point under Defense Access.
Read at SiliconANGLE ↗ • Read at Anthropic ↗
French AI company Mistral released Mistral Large 4, a 1 trillion parameter model nicknamed Le Chonk, through a guarded public preview and plans to publish its weights, the files that let anyone run the model, in about three weeks after safety testing, per TechCrunch. Vice President of Science Pierre Stock said Mistral will work with trusted partners and governments in the meantime to keep the weights from being used for malicious attacks. He said the model was trained on 4,000 Nvidia GPUs, two to three times fewer than its Chinese competitors use. Mistral calls Le Chonk the most capable open weight modelOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released. built outside China and says it trained the model from scratch, WIRED reported. The U.S. government has accused Chinese developers of distillationDistillationTraining a smaller or cheaper model on the outputs of a larger one so it inherits much of the larger model's ability. It is a standard technique, but it has become a trade dispute because it lets a lab reproduce a rival's capabilities without matching its compute spend., training a smaller model on a larger one's outputs.
Read at TechCrunch ↗ • Read at WIRED ↗ • Read at CNBC ↗
Reflection unveiled Beam on Monday, a model with 501 billion parameters built for coding, reasoning and agent work, AI Business reported. The company said it trained Beam on 23.8 trillion tokens, or word fragments, drawn from online and licensed proprietary data. Reflection positions Beam as comparable to GLM-5.2 from Chinese developer Z.ai. Nvidia has put $800 million of equity into Reflection and gives it access to its chips. A separate deal for computing capacity at SpaceXAI's Colossus data center costs Reflection $150 million a month through 2029.
Read at AI Business ↗ • Read at Reflection ↗
IBM and its Red Hat unit said engineers in Lightwell, their open source security program, have repaired more than 400 vulnerabilities nobody had previously identified in popular Java libraries, according to SiliconANGLE. The companies tied the work to AI agents' growing ability to chain several low risk software weaknesses into one serious attack. Engineers built the fixes for the older library versions that businesses still run in production. The companies also made Lightwell Clearinghouse, a service where enterprise customers can submit specific open source components they depend on for priority repair, generally available. Gunnar Hellekson, who runs Lightwell at Red Hat, said AI agents now go after old dependencies at machine speed.
Read at SiliconANGLE ↗