Rosetta Intel

Product and model updates from OpenAI / Anthropic / Google. 🛡 = security-related.

OpenAI

Safety Bug Bounty ($1M annual pool, operated via Bugcrowd)remains live — scope covers prompt injection, guardrail jailbreaks, training-data leakage, DoS, and the OWASP LLM Top 10, paying $500–$15,000 per finding. One of the most worthwhile bounty channels for AI security researchers right now.
Wraith — AI Bug Bounty Programs
The open-weight gpt-ossseries continues in the product lineup: 120b (runs on a single GPU, performance close to o3/o4-mini) and the lightweight 20b (runs locally on high-end consumer laptops).
LLM Stats — AI Updates

Claude Fable 5(released 2026-06-10) targets the broader product line; Claude Mythosremains limited to Glasswing partners.
Superhuman — New Models
Project Glasswing(the Claude Mythos security program): Chinese-language community reports say it has already surfaced 10,000+ high-severity vulnerabilities for 40+ organizations — model-driven vulnerability hunting is entering the commercial stage.
FreeBuf (via CSDN)
Compute: locked up the entire capacity of SpaceX/xAI's Colossus 1data center (Tennessee) — 300+ MW, 220,000+ Nvidia GPUs.
QverLabs — 2026 AI Arms Race

Gemini 3.5 Proteased for June delivery (Pichai: "give us until next month"); Gemini 3.5 Flashis already GA and now the default for the Gemini app and Search AI Mode.
llm-stats
AlphaEvolve(the Gemini-powered coding agent) expands into genomics, quantum physics, mathematics, and AI infrastructure — DNA sequencing error detection down 30%, quantum-circuit error down 10x.
QverLabs
Genie 3world model: can generate interactive 3D environments lasting several minutes at 720p/24fps.
WaveSpeed — June 2026 AI Launch Wave