Rosetta Intel

OpenAI / Anthropic / Google daily updates · with a focus on security-related items (🛡)

OpenAI

No major product/model releases over the weekend. GPT-5.5 Instantremains ChatGPT's default model (better accuracy and web-search use). Having completed its confidential IPO filing, OpenAI continues jockeying for position with Anthropic, which filed first.
Ecosystem: Nvidia previously announced OpenAI as one of the first customers of its next-generation Verachip.

Project Glasswing expands to about 150 organizations, opening the Claude Mythoscybersecurity preview. Anthropic says early partners have already used it to find 10,000+ high-severity vulnerabilities— a landmark data point for AI applied at scale on the defensive side.
- Background: Mythos is the same model as the public Fable but with the cyber safety guardrails removed — a restricted version open only to vetted parties (via Project Glasswing).
Capital: previously filed for IPO(valuation of roughly $965B, annualized revenue of about $47B).

Gemini Spark— a general-purpose AI agent that can reason across connected apps — rolls out in beta to trusted testers and AI Ultrasubscribers.
"Omni" world modelreleased.
Gemini 3.5 Flashremains the default model in the Gemini app and Search AI Mode (GA at Google I/O on 5/19; API $1.50 / $9.00 per million tokens).
Lifecycle: two Gemini preview models will be shut down on June 25— migration required.

Anthropic Project Glasswing / Claude Mythos: AI-assisted vulnerability discovery at scale (10,000+ high-severity findings). Defensive capability is spreading — which also implies potential spillover risk of equivalent offensive capability.
Google Gemini Spark (agent): once a general-purpose agent connects to "connected apps", the prompt-injection → tool-execution attack surface expands accordingly (see the main brief's Web/AI security section, "When prompts become shells", and the research on GPT-5/Gemini's 79% direct-injection failure rate).