- Multi-agent workbench built on Gemini 3.1 — a project coordinator fans research questions out to sub-agents that write code, search literature, and attempt proofs.
- Solved 23 of 48 problems (48%) on Epoch AI's FrontierMath Tier 4, more than double the 19% scored by the underlying Gemini 3.1 Pro and a new SOTA.
- Oxford's Marc Lackenby cracked Problem 21.10 from the Kourovka Notebook after a reviewer agent flagged a flaw in the AI's first proof attempt and surfaced a strategy he could finish.
- Stateful, asynchronous workspace that tracks failed hypotheses and outputs LaTeX with margin annotations — DeepMind frames it as 'Claude Code for mathematicians.'
- New cybersecurity platform that builds an editable threat model from your repo, validates exploitable vulnerabilities in isolated sandboxes, and drafts patches with audit-ready evidence.
- Built on three models: stock GPT-5.5, GPT-5.5 with Trusted Access for Cyber for verified defenders, and GPT-5.5-Cyber — a permissive variant for red-team and pentest work.
- Launch partners include Cloudflare, Cisco, CrowdStrike, Palo Alto Networks, Oracle, Akamai, Fortinet, Zscaler, Snyk, Semgrep, and Socket.
- Reads as OpenAI's direct response to Anthropic's Claude Mythos and Project Glasswing; available by request, broader rollout planned with industry and government partners.
- Akamai's largest contract ever — roughly 6% of its annual revenue once fully ramped — to host Claude inference on Akamai's edge cloud.
- Anthropic CEO Dario Amodei cited 80x year-on-year growth in annualized revenue and usage in Q1; first economic impact expected by end of 2026.
- Adds to Anthropic's compute scramble alongside SpaceX Colossus 1 and the $100B+ AWS commitment.
- Akamai stock closed Friday up 27% at $148.38 on the news.
- Open-OSS/privacy-filter typosquatted OpenAI's real release, copied the model card verbatim, and shipped a loader.py that pulled a Rust info-stealer for Windows.
- Payload disables AMSI and ETW, runs VM and debugger checks, and runs eight parallel collectors targeting Chromium, Firefox, Discord, and crypto wallets.
- HiddenLayer telemetry linked six more malicious repos under the same account, all uploaded April 24 — likely a coordinated supply-chain campaign.
- Spent days on Hugging Face Trending before removal; covered by BleepingComputer, The Hacker News, and Security Boulevard.
- ICML 2026 paper introduces TwELL (Tile-wise ELLPACK), a sparse packing format that materializes inside the gate-projection epilogue — no extra kernel launch, no extra memory read.
- Custom CUDA kernels deliver 20.5% inference and 21.9% training speedups on H100s with no accuracy loss versus dense baselines.
- Reference training code and kernels released open source on GitHub (SakanaAI/sparser-faster-llms).
- Aimed at making unstructured sparsity a practical efficiency lever for frontier-scale models.
- Draft executive order would expand federal cybersecurity information-sharing programs to include AI companies — but stops short of requiring government approval of frontier models.
- Directs agencies to partner with labs on AI-enabled cyber defense; framed by officials as a response to tools like Anthropic's Mythos finding decades-old vulnerabilities.
- WaPo: national security agencies are pushing for more sway over AI rules; Commerce is resisting any pre-release testing mandate.
- Lands the same week as OpenAI's Daybreak launch and Google's disclosure of an AI-built zero-day campaign.
01
DeepMind's AI Co-Mathematician Sets a New High on FrontierMath Tier 4
research arxiv.org
02
OpenAI Launches Daybreak to Plant GPT-5.5 Inside Software Security Pipelines
tools openai.com
03
Anthropic Signs $1.8B, Seven-Year Compute Deal With Akamai
industry bloomberg.com
04
Fake OpenAI Privacy Filter Repo Hits #1 on Hugging Face With 244K Downloads
community hiddenlayer.com
05
Sakana AI and NVIDIA Open-Source TwELL Sparse Kernels for 20%+ LLM Speedup
open-source sakana.ai
06
Trump Drafts AI Security Order That Skips Mandatory Model Tests
industry washingtonpost.com