AI News

GPT-5.6-Cyber answered 95% of exploit-chain, auth-bypass, and privilege-escalation queries in internal tests, versus 1.5% for the standard GPT-5.6-Sol — the gap is deliberate. Daybreak now runs two tiers: Blue (frontier general-purpose models for everyday security work) and Red (the Cyber model, gated behind tighter vetting for vulnerability research and penetration testing).

A distilled version of Muse Spark 1.2, Glimmer targets local agentic workflows — scheduling, document processing, coding assistance, and LLM-as-judge tasks — without sending data to cloud APIs. Weights are live on Hugging Face; a more capable Muse Spark 1.2 cloud update is also planned. This is a distinct release from the Muse Code terminal agent Meta launched last week.

The offering includes a 30-day underwriter option for an additional $2.25B, pushing potential proceeds to $17.25B, and lifts Intel’s 2026 capex target above $20B. Intel is explicitly targeting physical AI, purpose-built silicon, and advanced packaging as growth vectors — a marked turnaround for a company that was trailing Nvidia and AMD just a year ago.

Founded in 2023 by a former AMD director, Taalas customizes two metal layers of a chip per model and turns around model-specific silicon via TSMC’s 6nm process in about two months — a structurally different approach to inference efficiency than software-level optimization. Financial terms were not disclosed; regulatory close is expected in Q4 2026.

Article 50 transparency obligations are now in force across all EU member states; fines run up to €15M or 3% of worldwide annual revenue. A separate European Parliament vote pushed the high-risk AI system compliance deadline back to December 2027, but providers serving EU users — regardless of where they’re incorporated — must meet the transparency rules now.

Like the first deal, the new package would route through a special-purpose vehicle that buys Google’s custom chips and leases them to Anthropic, keeping the debt off Anthropic’s balance sheet. Terms are not finalized — but with $71B in private credit potentially stacked against a single AI lab’s compute needs, this marks a new scale of infrastructure financing in the industry.