Rosetta Intel

Daily developments from OpenAI / Anthropic / Google. 🛡 = security-related. Window: past 24 hours (2026-06-11 → 2026-06-12).

OpenAI

Filed confidentially for an IPO(around 6/8), roughly a week behind Anthropic; the last valuation was about $852B(post-money). The broader AI IPO wave (including the 6/11 SpaceX pricing signal) keeps heating up.
ChatGPT "Lockdown Mode"— restricts tools and connectors that could be abused via prompt injectionfor data exfiltration in browsing/agent scenarios. A continuation of the response path OpenAI set after acknowledging that browser injection "may never be fully fixable."
THN — Lockdown Mode

Claude Fable 5 / Mythos 5 split (GA 6/9)— an unusual "one model, two brands" release: the same underlying model split by a layer of safety classifiers.
- Fable 5(for the public): requests that trip cyber/bio/chemical/distillation classifiers are routed down to the weaker Opus 4.8; available via the Claude API and included in Pro/Max/Team/Enterprise.
- Mythos 5(restricted): the same model with cybersecurity guardrails removed, open only to vetted cyber defenders and critical infrastructure operators, which Anthropic calls "the world's strongest cybersecurity model."
- Pricing: both at $10 / $50per million input/output tokens, less than half the earlier Mythos Preview.
  THN — Fable 5 / Mythos 5
Funding: closed a new round at a $965Bvaluation (surpassing OpenAI), and has filed confidentially for an IPO.

Gemini 3.5 Prois expected "this month" (at I/O it was "next month"; still no firm date).
Gemini 3.5 Flash(announced at I/O on 5/19) is now the default model for the Gemini App and Search AI Mode.

Anthropic's Fable/Mythos safety classifier tieringis itself a governance experiment in "distribute by capability, admit by credential," and security teams should watch its impact on regulated industries.
OpenAI's Lockdown Modeechoes this week's AI security throughline (Langflow RCE in the wild, LLM agents hitting Salesforce, Teams voice cloning): frontier vendors are tightening agent/tool permissions on the product side while the offensive side weaponizes agents in parallel.

Configuration is in the ai_frontier and ai_security sections of intel/sources.yaml.