AI Daily Dev — September 16, 2026

Google Opens Claude Opus 5 to All Its Engineers Through Antigravity

Sept 15 Business Insider scoop: Google reverses its Gemini-only internal rule and turns on Anthropic's Opus 5 inside Antigravity, its internal build of the Antigravity IDE.
Access is company-wide with per-user quotas; standalone Claude Code and OpenAI Codex still blocked — engineers must pick the Opus 5 model inside Antigravity.
Google's line: 'Gemini remains our primary and foundational model for internal development' — third-party models fill 'specialized use cases.'
Reads as a public concession from a $40B Anthropic investor that Claude wins on coding today; Techmeme leading item, HN comments running on rival-lab optics.

industry businessinsider.com

Sept 15 flip: Cloudflare's default settings now block AI crawlers on any ad-supported page unless the operator declares a distinct search-only vs. training/agent bot.
Applies to all new customers, new sites on existing customers, and every existing free-tier account; Cloudflare fronts 20%+ of web domains.
Pay Per Crawl becomes Pay Per Use: publishers get paid when their content shows up in AI answers, not just when a bot fetches. Launch partners Ceramic.ai and You.com.
Agent builders scrambling — mixed-use bots that don't split identities lose access; expect ranking and citation shakeups over the next week.

tools cloudflare.com

Sept 11: Atria-Dawn-Preview appears on Hugging Face with no press release; Sept 14 a 143-author arXiv paper follows.
744B-parameter agentic MoE built on GLM-5.2, tuned for the full analyze → design → tool-use → code → run → verify loop.
Code and weights MIT-licensed — any GPU-rich team can deploy a frontier open agent with no licensing negotiation.
Evaluated on 16 benchmarks plus 769 real task records from 56 users; ~1/3 of completed tasks rated infeasible without the model.

open-source huggingface.co

Sept 9 GreyNoise/Blackpoint Cyber report: single actor built exploits for CVE-2026-81578 and CVE-2026-82078 in PaperCut NG/MF, then let an agent swarm run the campaign.
440 compromised instances across 395 orgs in 48 countries; 11 orgs breached in 26 seconds; RCE on first victim under 4 hours from empty workspace.
Harness stack: OpenAI Codex driving loops, a DeepSeek model, and off-the-shelf offensive tools; credentials from 280, OS/domain secrets from 147, admin at 12.
Concrete data point for the 'lone actor = state-team capability' claim Anthropic and CISA have been making for two weeks.

industry thehackernews.com

Sept 14: Intel spinoff Cornelis Networks raises $205M led by IAG Capital and launches Active Compute Fabric with Qualcomm.
Design does in-fabric compute so accelerators keep working while data moves — targets the GPU-idle-waiting-for-data waste that eats real-world inference throughput.
CN5000 400Gbps switch shipping now; CN6000 800Gbps sampling with customers, wider availability Q4 2026.
First credible non-Nvidia scale-up fabric with a top-tier accelerator partner attached.

industry techcrunch.com

Sept 14–15: White House AI/Crypto advisor David Sacks calls the 'pace the frontier' push a bid for regulatory capture, not a safety measure.
Argues the two frontier labs can unilaterally delay any model they think is unsafe; rules would only pin down competitors chasing them.
Rejects the antitrust waiver OpenAI's policy chief said the labs don't need — and points to product liability and market pressure as sufficient discipline.
HN discussion (id=49685991) mixed but skeptical of coordination optics; frames the Bloomberg 'weeks of safety talks' story as an incumbents-only club.

industry hackernews