Skip to content
Source · Daily Brief

AI Daily Brief — 30 November 2025

Sunday closed November on two safety/policy shockwaves. Icaro Lab “Adversarial Poetry” paper hit mainstream coverage — hand-crafted poems achieved 62% jailbreak success on average across 25 frontier models from OpenAI + Google + Anthropic + Meta + xAI + DeepSeek + Qwen + Mistral + Moonshot; some providers exceeded 90%. Meta-prompt verse conversions of MLCommons prompts produced ASRs up to 18× baseline. Google Gemini + DeepSeek + MistralAI “consistently provided answers” under poetic prompts; OpenAI GPT-5 + Anthropic Claude Haiku 4.5 showed strongest resistance. Attack surface spanned CBRN + manipulation + privacy + misinformation + cyberattack facilitation. Authors declined to publish full poem corpus, calling it “too dangerous to share.” OpenAI “Code Red” framing crystallizes publicly by Sunday — reporting confirmed Sam Altman issued internal memo in response to Gemini 3 momentum, directing OpenAI to deprioritize AGI research and Sora 2 video app and focus on shipping ChatGPT upgrade. Pivot would land days later as GPT-5.2. Weekend marked inflection point in public framing of OpenAI-Google rivalry. Kuala Lumpur Parliamentary Conference on Responsible AI wrapped — joint Parliament of Malaysia / Commonwealth Parliamentary Association / IPU / UNDP gathering Nov 28-30. Brought together legislators from Commonwealth + IPU member states to coordinate national AI legislation approaches — quieter multilateral counterpoint to US-China-EU axis. DeepSeek Math-V2 weekend uptake spikes — most-discussed open model of weekend. Coverage focused on self-verification (model proves own derivations before answering) + DeepSeek delivering IMO-gold parity with permissive license that US labs only achieved with closed weights months earlier. Brookfield AI Infrastructure Fund reverberates as weekend deal-flow story — $10B target BAIIF (launched Nov 19 with $5B from Brookfield + NVIDIA + Kuwait Investment Authority); plans up to $100B across energy + land + DCs + compute; seed $5B framework with Bloom Energy for up to 1 GW behind-the-meter DC power.

Top stories

  • Adversarial Poetry: 62% jailbreak success across 25 frontier models. Some providers >90%. 18× ASR uplift via MLCommons verse conversion. via Engadget
  • Gemini + DeepSeek + Mistral most permissive; GPT-5 + Haiku 4.5 most resistant. Attack surface: CBRN + manipulation + privacy + misinformation + cyber. Full poem corpus withheld as “too dangerous.” via arXiv
  • Altman “Code Red” crystallizes publicly. Internal memo: deprioritize AGI + Sora 2 video app; focus on ChatGPT upgrade. Response to Gemini 3 momentum. Pivot lands days later as GPT-5.2. via TechCrunch
  • Kuala Lumpur Parliamentary Conference on Responsible AI wraps. Parliament of Malaysia + CPA + IPU + UNDP. Nov 28-30. Commonwealth + IPU legislators coordinating national AI legislation. via IPU
  • DeepSeek Math-V2 weekend uptake spikes. Most-discussed open model. Self-verification + IMO gold + permissive license. US labs only achieved this with closed weights months earlier. via ByteIota
  • Brookfield AI Infrastructure Fund reverberates. $10B target BAIIF; $5B from Brookfield + NVIDIA + KIA. Up to $100B across energy + land + DC + compute. $5B Bloom Energy framework for up to 1 GW behind-the-meter DC power. via Brookfield

Who shipped

Nobody Sunday. OpenAI, Anthropic, Google, Meta, xAI, DeepSeek, Alibaba dark. Adversarial Poetry paper + Code Red coverage + DeepSeek Math-V2 echo dominate.

Open-source pulse

Math-V2 continued downloading + benchmarking; first open model to reach IMO gold. November closed with Chinese open-weight ecosystem at structural advantage — Qwen + DeepSeek + Kimi K2 + Math-V2 + Hunyuan-1.5 + ERNIE 5.0.

Money, infra & hardware

Brookfield’s $100B AI infra program is the structural weekend signal — institutional capital scale at hyperscaler levels. November funding momentum (Cursor $2.3B + Prometheus $6.2B + Lambda $1.5B) continues to reset AI valuation baselines.

Quiet corners

Adversarial Poetry is the year’s most consequential safety-research result — single-turn jailbreak via verse cracks 62% of frontier models. Authors withholding corpus signals tier of severity.

By the numbers

  • 62% / >90% / 18× — Adversarial Poetry avg success / worst-case provider / ASR uplift via MLCommons conversion
  • 25 — frontier models tested
  • Nov 28-30 / 3 days — Kuala Lumpur Parliamentary Conference
  • $10B / $5B / $100B / 1 GW — Brookfield BAIIF target / committed / max deployment / Bloom power
  • Most-mentioned topic: Adversarial Poetry jailbreaks
  • Quietest segment: fresh frontier launches

Compiled by AI Feed’s editor from verified web sources for 30 November 2025.