Anthropic open-sourced its defending-code-reference-harness for AI-powered vulnerability discovery — the most concrete first-party release of frontier-lab security tooling to date (443 HN points). Apple quietly approved Poke as the first AI agent on its Messages for Business platform, the first real third-party agent surface inside Apple's commerce stack ahead of WWDC. The AI-IPO race kept warming: Daniela Amodei pushed back on doubts about returns ahead of Anthropic's listing, and Wired documented investors backing both Anthropic and OpenAI rather than picking sides. Compute and physical-infra constraints sharpened — TSMC said it can only support so much demand, Meta started literally pitching tents over data centers to cut cost, and Kevin O'Leary halved his planned 40,000-acre Utah build under local pressure. The research wave centered on planning and values: AdaPlanBench tests agents under progressively revealed constraints, RobotValues asks whether household robots make good choices when human values conflict, and Meta-Cognitive Memory Policy Optimization sharpens the credit-assignment problem for long-horizon agents.
Anthropic open-sourced its defending-code-reference-harness for AI-powered vulnerability discovery — a first-party frontier-lab security tool. In parallel, Daniela Amodei pushed back on doubts about AI returns ahead of Anthropic's IPO, and Wired's reporting shows top investors backing both OpenAI and Anthropic rather than picking sides.
Anthropic open-sources defending-code-reference-harness — a reference framework for AI-powered vulnerability discovery, drawing 443 HN points.
Anthropic president Daniela Amodei publicly counters skeptical-returns narratives — preparing the IPO market for breakneck-growth disclosures.
Wired finds top venture investors are taking positions in both OpenAI and Anthropic — 'why wouldn't you want to be in both Pepsi and Coke?'
Thinking Machines' Mira Murati re-engages publicly — the 'remain heads-down' window is closing as the cycle intensifies.
Apple approved Poke as the first AI agent on its Messages for Business platform — the first sanctioned third-party agent inside Apple's commerce stack — days before a WWDC widely expected to launch a Siri revamp. Apple also disclosed $1.4T in App Store billings, and reportedly weighing cameras for next-gen AirPods. Apple is moving from late to credible on agents.
Poke becomes the first sanctioned third-party AI agent inside Apple's Messages for Business — a real opening of Apple's commerce surface to outside agents.
WWDC 2026 preview: Siri revamp expected to land, Apple Intelligence updates across the OS lineup.
Apple discloses $1.4T in App Store billings (up from $1.3T), with 90% commission-free — useful framing ahead of antitrust and AI-distribution arguments.
Wired examines the rumored next-gen AirPods with cameras — a possible Apple wedge into ambient AI hardware.
TSMC said it can only support so much AI demand, Meta is pitching tents over data centers to cut build cost, and Kevin O'Leary halved his planned 40,000-acre Utah build under local pressure. The supply ceiling and the local-opposition ceiling are now both visibly binding.
TSMC publicly acknowledges it cannot keep up with AI demand — the binding constraint on frontier silicon now spoken aloud.
Meta is using tent structures over data-center buildouts to slash construction time and cost — Tesla-style improvisation.
Kevin O'Leary halved his planned 40,000-acre Utah data center under sustained local opposition.
Ars details operator-side cooling and water-recovery efforts as scrutiny over AI's water footprint intensifies.
MIT Technology Review frames the Meta Instagram exploit as proof that AI-security needs more than Mythos-style frontier-lab programs. Nemotron 3.5 Content Safety from NVIDIA, Open Code Review from Alibaba, and an Estonian benchmark on resisting Russian propaganda all add tooling and measurement to the layer.
MIT Tech Review frames the Meta AI support-agent Instagram exploit as evidence that frontier-lab safety programs (Mythos, Daybreak, Glasswing) are necessary but not sufficient.
NVIDIA ships Nemotron 3.5 Content Safety — customizable multimodal content-safety models for enterprise deployments.
Alibaba open-sources an AI-powered code-review CLI tool (191 HN points).
Estonian government benchmark scores dozens of models on resistance to Russia's strategic narratives — first state-issued geopolitical-bias eval.
AdaPlanBench measures adaptive planning under progressively revealed user and world constraints, RobotValues evaluates household robots when human values conflict, Dream.exe asks whether video-gen models can actually drive executable robot manipulation, and Meta-Cognitive Memory Policy Optimization sharpens long-horizon credit assignment.
Benchmark for adaptive planning where both world and user constraints emerge through interaction — closer to real customer-facing agent work.
Evaluates household robots on actions when values conflict — task success vs autonomy, efficiency vs social appropriateness.
Tests whether video generative outputs can be turned into executable robot manipulation rather than just plausible footage.
Trains agent memory policies with meta-cognitive credit assignment — localizes where intermediate memory quality matters.