OpenAI's enterprise + cyber-defense push
OpenAI unveiled a two-pronged enterprise expansion: the OpenAI Deployment Company, a majority-owned vehicle launching with 150 Forward Deployed Engineers and Deployment Specialists plus $4B from 19 investment, consulting, and SI partners (@sama, @openai), powered in part by an acqui-hire of Tomoro (@openai). The pitch is helping enterprises get frontier AI into production — though Ethan Mollick was sharply skeptical, arguing the existence of dedicated "forward deployed" consulting groups is itself evidence the labs don't actually believe in near-term ASI (@emollick).
The second prong, Daybreak, pairs GPT-5.5 and Codex with security partners — Akamai, Cisco, Cloudflare among them — to automate vulnerability discovery, patch validation, and threat modeling (@sama, @openai, @thehackersnews). Altman framed it as getting in front of AI that is "about to get super good at cybersecurity" (@sama).
The backdrop is harsh: a Reddit thread cataloging GPT-5.5's pricing at $5/$30 per million tokens — roughly 2x GPT-5.4 — fueled questions about whether the "all-you-can-eat" subscription model is fraying just as OpenAI leans harder on enterprise services (last30days, reddit.com).
AI-amplified cybersecurity threats and breaches
The day's most striking data point: average time from CVE disclosure to working exploit has collapsed to ~10 hours in 2026 from 56 days in 2024, with AI-assisted attackers breaching systems in 73 seconds (@thehackersnews). Google reported the first confirmed AI-generated zero-day — a 2FA bypass on a popular open-source admin tool — caught pre-mass-exploitation, with telltale LLM artifacts including a fabricated CVSS score and over-commented code (@thehackersnews).
Active campaigns piled up: 2,000+ IPs exploiting cPanel CVE-2026-41940 to drop the Filemanager backdoor tied to Mr_Rot13 (@thehackersnews); a malicious Checkmarx Jenkins AST plugin published after TeamPCP allegedly breached the upstream repo, extending a streak that already hit KICS, VS Code extensions, and Bitwarden CLI (@thehackersnews); Silver Fox's Rust stealer pulling payloads from a JSON Keeper site (@thehackersnews); and Instructure paying ShinyHunters to halt leakage of 3.65TB of Canvas data from ~9,000 schools (@thehackersnews). "Bleeding Llama" — a malicious GGUF that makes Ollama read past memory — can exfiltrate API keys and Claude tool data (@thehackersnews). Apple's iOS 26.5 brought default E2EE to RCS as a rare bright spot (@thehackersnews). Last-30-days context: a Reddit writeup of Claude Code / Copilot / Codex compromises stressed that the real prize for attackers is user credentials, not the models (last30days, reddit.com).
Claude Platform on AWS and Claude Code agentic upgrades
Anthropic made the Claude Platform generally available on AWS, exposing Managed Agents, prompt caching, pre-built tools, and skills behind AWS IAM/billing, with same-day feature parity with the native API (@claudedevs, @openrouter). The Hacker News launch thread hit 151 points (last30days, claude.com).
Claude Code 2.1.139 added /goal, letting Claude work across turns toward a user-set completion condition in interactive, -p, and Remote Control modes (@steipete, @alexfinn), plus an agent-view research preview for managing parallel sessions (@bcherny, @claudedevs). Boris Cherny said Cowork on Opus 4.7 one-shot-booked 8 flights and 5 hotels (@bcherny). Gary Marcus, of all people, called Claude Code "the most neurosymbolic thing I have ever seen" — 53 tools, 500K lines of symbolic code wrapping an LLM — framing it as vindication, not victory, for pure LLMs (@garymarcus). Caveat from last-30-days chatter: Opus 4.7 reportedly regressed on MRCR v2 long-context retrieval at 1M tokens (78.3% → 32.2%) versus 4.6 (last30days, reddit.com).
Real-time interaction models, voice and visual intelligence
Thinking Machines introduced TML-Interaction-Small (276B-A12B), a model trained from scratch for native real-time interaction rather than retrofitting onto turn-based architectures — pitched as solving the human↔AI bandwidth bottleneck (@swyx). OpenAI shipped gpt-realtime-2, processing speech natively rather than transcribing, with Mollick noting it's much smarter than the GPT-4o-era predecessor but will force prompt rewrites (@emollick). Karpathy argued vision is the right output channel for AIs given how much human cortex is dedicated to it, recommending "structure your response as HTML" as a daily-driver trick (@karpathy). Fei-Fei Li reiterated that language-model fixation misses the physical, perceptual economy (@garymarcus), and BFL teased models that understand motion and interaction, not just images (@aidotengineer).
AI scrutiny: hype, hallucinations and OpenAI investigations
A new Lancet paper found fabricated citations in biomedical papers have grown 12x since 2023 (@emollick, @garymarcus). Rep. James Comer's House Oversight Committee formally requested information from Sam Altman about financial conflicts, citing an undisclosed OpenAI board audit committee (@garymarcus). Ilya Sutskever reportedly has a dossier on Altman's alleged dishonesty (@garymarcus). Marcus also cited an NBER paper finding 9-in-10 executives report no employment or productivity impact from AI in 3 years, and accused Anthropic of overhyping its "Mythos" model as a cyberweapon (@garymarcus). Mollick lamented 3,000-word AI-slop posts from high-status tech accounts pulling 1M+ views (@emollick).
Open-weights, ML acceleration, benchmarks and dev tools
Clément Delangue showed local open-weight intelligence outpacing Moore's Law: on the same 128GB MacBook Pro, top runnable open models jumped from Llama 3 70B (score 10) to a Q2 DeepSeek V4 Flash (47) on the Artificial Analysis index over 24 months (@clementdelangue). OpenRouter shipped Pareto Code, a router that picks the cheapest model clearing a quality bar (with a :nitro throughput variant), currently topped by DeepSeek V4 Pro, GPT-5.4 Mini, and Gemini 3.1 Pro (@openrouter). Ring-2.6-1T is free on OpenRouter through May 15 (@openrouter). ml-intern logged 1M messages in 3 weeks — "3.3 agent-years of ML research" — including a from-scratch 100M DeepSeek V4 replication (@clementdelangue, @_akhaliq). Mollick's PACT negotiation benchmark put GPT-5.5 at #1 (@emollick). DeepMind + Sainsbury Lab used AlphaFold's Structural Novelty Index to find an 11-protomer complex (@googledeepmind), NVIDIA pushed Jetson Orin and Vera Rubin Space-1 into orbit (@nvidia), and Consensus raised $30M for an "AI OS for Research" (@swyx).
The Bottom Line
Today crystallized AI as enterprise infrastructure and as attack surface in the same breath: OpenAI built a deployment company and Anthropic shipped on AWS while AI-generated zero-days, supply-chain compromises, and 10-hour CVE-to-exploit windows arrived in production. Underneath, real-time interaction models, neurosymbolic agents, and laptop-class open weights kept compounding — even as a Lancet citation crisis, a House probe of Altman, and NBER productivity data forced harder questions about whether the hype matches the receipts.