Anthropic team warns of extinction; Christiano joins OpenAI — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Evan Hubinger posted on X that there is a greater than 10% chance AI 'could kill all humans' within the next decade and that Anthropic has 'no plan to solve alignment for superintelligence.'link ›
- Jacob Coxon, who resigned from Anthropic after stints at OpenAI, wrote 'neither company is acting responsibly' and warned superhuman systems that can 'hack anything' are imminent.link ›
- Anthropic withheld its latest model from the UK's AI Security Institute, the Financial Times reported; a Cabinet Office spokesperson declined to confirm or deny.link ›
- Anthropic disclosed it disrupted a Yemen-based, Iran-linked guided-weapons cell that used Claude to build missile and rocket guidance software, per a Bloomberg threat intelligence report.link ›
- Paul Christiano joined the OpenAI Foundation board and its Safety and Security Committee, giving a vocal 'loss-of-control' critic a direct governance seat at the IPO-bound lab.link ›
- Tristan Buckmaster alleged OpenAI learned of his work with Levent Alpöge on Navier-Stokes and deployed an internal model to solve the Millennium Prize, comparing it to Deep Blue vs. Kasparov.link ›
- Meta launched Muse, a personal AI agent handling shopping, email, and trip planning, with Stripe Link checkout and an isolated cloud VM watched by a Sentinel oversight agent.link ›
- ON.energy reframed the AI-grid debate after a July 22 Ashburn transmission fault dropped more than 3 GW of data center load, pitching medium-voltage inline UPS systems as the missing architecture.link ›
Anthropic alignment researcher Evan Hubinger posted on X that he believes there is a greater than 10% chance AI 'could kill all humans' within the next decade, adding the company has 'no plan to solve alignment for superintelligence.' The Financial Times separately reported Anthropic withheld its latest model from the UK's AI Security Institute — one of the only bodies tasked with independently assessing exactly that risk. Former researcher Jacob Coxon, who resigned from Anthropic after stints at OpenAI, wrote 'neither company is acting responsibly.' All of this lands weeks before Anthropic and OpenAI's anticipated IPOs, with UN adviser Dame Wendy Hall suggesting the public confessions might be 'PR and marketing' timed to the listing. The frontier labs aren't hiding the existential risk from investors — they're disclosing it first.
The stories behind this week
3 GW Virginia AI outage was an architecture failureON.energy reframes the AI-grid debate from generation (more turbines, solar, transmission) to the UPS architecture inside the data center fence — a spending shift toward medium-voltage inline systems if hyperscalers and utilities adopt the model. For utilities, certifying one medium-voltage box per site instead of every component could compress permitting timelines on the next gigawatt-scale wave, while turning a 70% load-swing liability into a grid asset that earns money in demand response.
Besxar Builds Orbital Chip Factory on SpaceX RocketsBy using existing rocket flights instead of building costly ground-based clean rooms, Besxar reduces upfront infrastructure barriers to advanced chipmaking. The approach could shift how semiconductor capacity is developed, especially if orbital manufacturing proves scalable and cheaper than fighting Earth’s physics.
Meta Launches Muse AI Agent for Everyday TasksMeta's WhatsApp and Instagram reach gives Muse a distribution path no rival personal AI agent currently matches, but the company's documented history of AI missteps — from the Discover data leak to a chatbot that enabled 20,000-plus Instagram account takeovers — makes consumer trust the real product risk, and Sentinel's design acknowledges that an agent handling payments and passwords demands infrastructure competitors haven't had to build for consumer audiences.
Anthropic Disrupts Yemen Cell Using Claude for Missile GuidanceAnthropic's own report places Claude inside an active weapons-development pipeline operated by an Iran-linked militant cell, forcing the company to defend its safety claims against evidence it published itself. The coverage flags that frontier models are no longer hypothetical proliferation risks.
Anthropic Researcher: >10% Chance AI Kills All HumansTwo of Anthropic's own researchers are publicly warning of existential risk while the company simultaneously withholds its latest model from the UK body specifically tasked with assessing that risk. For investors weighing Anthropic and OpenAI's anticipated IPOs and for UK regulators expecting AISI access, the gap between public safety warnings and disclosed models is now a live governance question with no public timeline for resolution.
Fidji Simo joins Nscale board, stays OpenAI adviserSimo is holding simultaneous roles at OpenAI (part-time adviser) and Nscale (board seat) — and Nscale builds AI infrastructure that directly serves the kind of compute needs OpenAI relies on, making the dual position one with clear strategic overlap.
Paul Christiano joins OpenAI Foundation boardChristiano's dual appointment to both the foundation board and the Safety and Security Committee gives a prominent AI-alignment advocate a direct governance role over OpenAI's safety posture — notable because he has publicly argued the industry is not on track to reduce acute loss-of-control risk to an acceptable level.
Mathematician: OpenAI Used My Work for Navier-StokesThe allegation reframes a celebrated research milestone as an alleged ethics violation, putting OpenAI's handling of user-submitted mathematical work under direct public scrutiny at the moment of its biggest claimed breakthrough.
Why it matters: Anthropic's own alignment team putting a 10%+ extinction probability on its product, while the company withholds that product from the UK's AISI, forces imminent IPO investors to underwrite existential risk without independent verification.
Ask SkimNews
